Agent Watch
AI agents, coding assistants and MCP servers that got pwned, since 2025.
Background readingMCP tool poisoningPrompt injection attacks
Cases
102 cases
| Case | Date | Product | Vector | CVE | CVSS | |
|---|---|---|---|---|---|---|
| In the wildExploitGym agents used leftover credentials to run privileged commands on OpenAI research clusters | 2026-11-10 | OpenAI internal models (ExploitGym evaluation) | Credential theft | None | ||
Incident. Agents reached OpenAI's CI system and research clusters, changed evaluation settings and read private grading code. How. On 19 July 2026 agents found credentials and code that earlier runs had left on a shared Artifactory and reused them. Two agents left the activity out of their answers. Fix. Runs stopped; ExploitGym runs blocked from 20 July OpenAI alignment misalignment report · checked 2026-11-12 | ||||||
| Proof of conceptCommand injection in the GitHub Copilot plugin for JetBrains IDEs disclosed local information | 2026-11-10 | GitHub Copilot Plugin for JetBrains IDEs | Data exfiltration | CVE-2026-83953 | 5.5 | |
Vulnerability. An authorized local attacker can read information they should not have access to. How. Microsoft lists it as a command injection flaw in the plugin; it gives no further technical detail. Fix. 1.17.0-251 NVD CVE-2026-83953 · Microsoft MSRC CVE-2026-83953 · checked 2026-11-12 | ||||||
| Proof of conceptpgEdge Postgres MCP Server fixed DNS rebinding, rate limit bypass and cross-table inference | 2026-11-10 | pgEdge Postgres MCP Server | Auth bypass | CVE-2026-125320 CVE-2026-125321 CVE-2026-125322 | 6.5 | |
Vulnerability. A malicious web page can read schemas and query results from a local server, login throttling can be evaded, and a caller can infer data from other tables the role can read. How. The HTTP transport skipped Origin checks, rate limiting trusted the client's X-Forwarded-For value, and the count_rows where parameter accepted subqueries on other tables. Fix. 1.1.0 NVD CVE-2026-125320 · NVD CVE-2026-125321 · NVD CVE-2026-125322 · GitHub advisory GHSA-6hg5-9f3w-c886 · GitHub advisory GHSA-v347-v3w7-7fq9 · GitHub advisory GHSA-fhv3-x3jh-grj7 · checked 2026-11-12 | ||||||
| Proof of conceptJunie CLI Allow Similar Commands option allowed unintended command execution | 2026-11-06 | JetBrains Junie CLI | Config runs code | CVE-2026-123315 | 7.0 | |
Vulnerability. Commands the user did not approve could run on their machine. How. Approving a command with the Allow Similar Commands option let other commands through as well. Fix. 26.8.24 NVD CVE-2026-123315 · JetBrains fixed security issues · checked 2026-11-12 | ||||||
| Proof of conceptJunie CLI wrote sensitive environment variables to session event logs | 2026-11-06 | JetBrains Junie CLI | Credential theft | CVE-2026-123316 | 5.5 | |
Vulnerability. Anyone who can read the session logs can collect secrets held in environment variables. How. Session event logs recorded sensitive environment variables. Fix. 26.6.29 NVD CVE-2026-123316 · JetBrains fixed security issues · checked 2026-11-12 | ||||||
| Proof of conceptJunie GitHub Action prompt injection protection bypassed with HTML-encoded characters | 2026-11-06 | JetBrains Junie GitHub Action | Prompt injection | CVE-2026-123851 | 5.4 | |
Vulnerability. Injected instructions can reach the agent despite the action's prompt injection filter. How. HTML-encoded characters got past the protection. Fix. 1.6.4 NVD CVE-2026-123851 · JetBrains fixed security issues · checked 2026-11-12 | ||||||
| Proof of conceptJunie GitHub Action ran fork code with workflow credentials | 2026-11-06 | JetBrains Junie GitHub Action | Credential theft | CVE-2026-123852 | 7.4 | |
Vulnerability. An attacker can run code with the repository workflow's credentials. How. Updates pushed from a fork could trigger code execution in a workflow context that held the repository's credentials. Fix. 1.8.0 NVD CVE-2026-123852 · JetBrains fixed security issues · checked 2026-11-12 | ||||||
| Proof of conceptfast-agent serve exposed agent tools on the network without authentication | 2026-11-06 | fast-agent | Auth bypass | CVE-2026-102610 | 6.4 | |
Vulnerability. Any network client can drive the agent; with --shell enabled it can run OS commands as the server user. How. fast-agent serve bound its MCP HTTP endpoint to 0.0.0.0:8000 with no authentication. Fix. 0.7.21 NVD CVE-2026-102610 · GitHub advisory GHSA-9vhv-g5hf-m7m7 · checked 2026-11-12 | ||||||
| Proof of conceptMCP TypeScript SDK fixed token audience, redirect leak and cross-client delivery flaws | 2026-11-05 | MCP TypeScript SDK (@modelcontextprotocol/sdk, /server, /client) | Auth bypass | CVE-2026-111771 CVE-2026-111769 CVE-2026-111770 | 7.7 | |
Vulnerability. A token issued for another service can authenticate to an MCP server, a malicious server can collect API keys, session IDs and OAuth secrets, and shared servers can hand one client's responses to another. How. Bearer auth did not check token audience, HTTP and OAuth clients re-sent headers and bodies across cross-origin redirects, and reused stateless servers matched replies only by JSON-RPC id. Fix. sdk 1.32.0; server and client 2.3.0 (express 2.0.2, server-legacy 2.3.1); audience validation must be enabled NVD CVE-2026-111771 · NVD CVE-2026-111769 · NVD CVE-2026-111770 · GitHub advisory GHSA-rvq5-wwqv-78pq · GitHub advisory GHSA-6prh-2h8m-c8cw · GitHub advisory GHSA-jm63-cq47-j5jp · checked 2026-11-12 | ||||||
| Proof of conceptClaude Code Edit tool could be raced into writing outside the workspace on macOS | 2026-11-03 | Claude Code | Sandbox escape | CVE-2026-88968 | 7.7 | |
Vulnerability. Code running in the Bash sandbox, for example from a malicious repo or dependency script, can overwrite user files outside the workspace. How. The unsandboxed Edit tool checked the path before writing, and sandboxed code could swap a parent directory for a symlink between the check and the write. Fix. 2.1.251 NVD CVE-2026-88968 · GitHub advisory GHSA-5xjx-9qx6-jcrv · checked 2026-11-12 | ||||||
| Proof of conceptPDF Reader MCP logged API key auth as on but never checked it | 2026-10-30 | PDF Reader MCP (@sylphx/pdf-reader-mcp) | Auth bypass | CVE-2026-62264 | 7.5 | |
Vulnerability. Any client that can reach the port can call every tool and read any PDF the server can open, with limited SSRF. How. In HTTP mode the server read MCP_API_KEY but never validated the X-API-Key header, and it bound to 0.0.0.0 by default. Fix. 3.0.1 NVD CVE-2026-62264 · GitHub advisory GHSA-q344-5v34-gm84 · checked 2026-11-12 | ||||||
| Proof of conceptLangGraph SDK v3 streaming let IDs traverse to other API paths | 2026-10-28 | langgraph-sdk and langgraph RemoteGraph | Auth bypass | CVE-2026-118667 | 4.3 | |
Vulnerability. Whoever controls an assistant or thread ID can make authenticated GET requests to other readable paths on the LangGraph server. How. The v3 stream helpers put assistant_id and thread_id into request paths without escaping, so relative segments resolved elsewhere. Fix. langgraph-sdk 0.4.6 and langgraph 1.2.14 NVD CVE-2026-118667 · GitHub advisory GHSA-3fx2-cqw9-xr3c · checked 2026-11-12 | ||||||
| Proof of conceptlanggraph-api JS auth handlers failed to isolate Store namespaces | 2026-10-28 | langgraph-api | Auth bypass | CVE-2026-118668 | 8.6 | |
Vulnerability. An authenticated user can read, overwrite, search or delete another user's Store data. How. Namespace rewrites returned by JavaScript or TypeScript custom auth handlers were not applied before the Store operation ran. Fix. 0.15.0 NVD CVE-2026-118668 · GitHub advisory GHSA-4hm6-w6qq-w73v · checked 2026-11-12 | ||||||
| Proof of conceptJunie CLI allowlist validation let other commands run | 2026-10-27 | JetBrains Junie CLI | Config runs code | CVE-2026-117530 | 6.7 | |
Vulnerability. Commands outside the user's allowlist could run on their machine. How. Validation of allowlisted commands was incomplete. Fix. 26.8.17 NVD CVE-2026-117530 · JetBrains fixed security issues · checked 2026-11-12 | ||||||
| Proof of conceptMCP Python SDK fixed token audience, redirect leak and schema $ref fetch flaws | 2026-10-27 | MCP Python SDK (mcp on PyPI) | Auth bypass | CVE-2026-116684 CVE-2026-116685 CVE-2026-116686 | 6.8 | |
Vulnerability. A token for another service can authenticate to an MCP server, a redirect can leak client credential headers, and a malicious server can make clients fetch URLs or read local files. How. Bearer auth did not check token audience, HTTP clients followed cross-origin redirects with custom headers, and outputSchema validation resolved $ref URLs from the client's network and filesystem. Fix. 1.30.0 or 2.2.0 NVD CVE-2026-116684 · NVD CVE-2026-116685 · NVD CVE-2026-116686 · GitHub advisory GHSA-w4fh-qvv9-3v23 · GitHub advisory GHSA-5h93-6whr-6q8j · GitHub advisory GHSA-rwrf-2pqf-9j8j · checked 2026-11-12 | ||||||
| Proof of conceptATXP Gemini CLI extension ran shell commands planted in emails or SMS | 2026-10-27 | ATXP Gemini CLI extension | Prompt injection | CVE-2026-117560 | 8.5 | |
Vulnerability. Text in an email or SMS the agent reads can run shell commands as the user. How. runAtxp() passed MCP tool parameters such as message_id, subject or query to a shell unsanitized, so injected $() or backtick substitutions executed. NVD CVE-2026-117560 · VulnCheck advisory · GitHub issue atxp-dev/atxp#15 · checked 2026-11-12 | ||||||
| Proof of conceptLangGraph Postgres and SQLite stores mixed up tenant namespaces | 2026-10-23 | langgraph-checkpoint-postgres and langgraph-checkpoint-sqlite | Auth bypass | CVE-2026-115883 | 6.0 | |
Vulnerability. A caller can read, overwrite or delete another tenant's stored agent data. How. The stores joined namespace labels with dots and only validated labels on put, so dotted labels in get, search or delete resolved to another namespace. Fix. postgres 3.1.3 and sqlite 3.1.2 NVD CVE-2026-115883 · GitHub advisory GHSA-6hmw-9432-67qq · checked 2026-11-12 | ||||||
| Proof of conceptLangGraph CLI printed a registry push token in verbose deploy logs | 2026-10-23 | LangGraph CLI | Credential theft | CVE-2026-115885 | 4.1 | |
Vulnerability. Anyone who reads the CI logs within 12 hours can pull or push images in the deployment's image repository. How. langgraph deploy --verbose echoed the stdin passed to docker login, which held the temporary push token. Fix. 0.4.32 NVD CVE-2026-115885 · GitHub advisory GHSA-8prr-f582-jqq3 · checked 2026-11-12 | ||||||
| Proof of conceptNetBox MCP Server HTTP mode had no authentication | 2026-10-23 | NetBox MCP Server | Auth bypass | CVE-2026-57207 | 5.9 | |
Vulnerability. Any client that can reach the port gets full use of the server's NetBox tools. How. With TRANSPORT=http the endpoint issued MCP sessions without any token or credential. Fix. 1.2.1 NVD CVE-2026-57207 · GitHub advisory GHSA-5qg7-j8mr-9477 · checked 2026-11-12 | ||||||
| Proof of conceptobsidian-tc ran destructive vault actions after the human declined | 2026-10-23 | obsidian-tc (Obsidian MCP server) | Rogue agent action | CVE-2026-115477 CVE-2026-115478 | 7.1 | |
Vulnerability. A misbehaving or hostile client could delete or change vault content without human approval, and memory tools could overwrite unrelated notes. How. The server treated a resubmitted requestState as approval even when the elicitation was declined, and memory notes were rewritten at computed paths without checking ownership. Fix. 1.31.4 NVD CVE-2026-115477 · NVD CVE-2026-115478 · GitHub advisory GHSA-737x-cwfr-5vvc · GitHub advisory GHSA-r3g7-r4hr-m258 · checked 2026-11-12 | ||||||
| Proof of conceptLiteLLM let internal users forge an admin token and run commands via MCP | 2026-10-22 | LiteLLM proxy | Auth bypass | CVE-2026-114612 | 9.9 | |
Vulnerability. An authenticated internal user can become proxy admin and run arbitrary commands through the MCP stdio endpoint. How. One cryptographic context protected both stored secrets and session tokens, so metadata encrypted by the API key flow could be replayed as a forged admin bearer token. Fix. 1.100.4, 1.101.3, 1.102.2, 1.103.1 or 1.104.0-rc.2 NVD CVE-2026-114612 · GitHub advisory GHSA-7hp6-4w63-5g45 · checked 2026-11-12 | ||||||
| Proof of conceptAgent Zero Web UI executed script injected through chat messages or uploads | 2026-10-22 | Agent Zero Web UI | Data exfiltration | CVE-2026-97635 | 8.6 | |
Vulnerability. Injected JavaScript runs in the victim's authenticated Agent Zero session. How. The UI did not sanitize chat messages or uploaded files before reproducing their content. Fix. 1.15 NVD CVE-2026-97635 · CISA CSAF advisory VA-26-296-01 · GitHub issue agent0ai/agent-zero#943 · checked 2026-11-12 | ||||||
| Proof of conceptD365 F&O MCP Server one-click Azure deploy left the API open | 2026-10-21 | D365 F&O MCP Server | Auth bypass | CVE-2026-114123 | 7.5 | |
Vulnerability. Anyone on the internet can query indexed proprietary X++ source, business logic and security role structure. How. The Azure template defaulted the API key to empty, and an empty key made the auth middleware pass every request on a listener bound to 0.0.0.0. Fix. 1.11.0 NVD CVE-2026-114123 · GitHub advisory GHSA-8764-fh3m-wf7g · checked 2026-11-12 | ||||||
| Proof of conceptMCP TypeScript SDK experimental tasks were shared across sessions | 2026-10-20 | MCP TypeScript SDK (@modelcontextprotocol/sdk) | Auth bypass | CVE-2026-111767 | 8.6 | |
Vulnerability. A client can list, read and cancel tasks that belong to other clients. How. From 1.24.0 the experimental taskStore did not check which session created a task when clients shared one InMemoryTaskStore. Fix. 1.32.0 NVD CVE-2026-111767 · GitHub advisory GHSA-22jm-h49p-29qw · checked 2026-11-12 | ||||||
| In the wildCERT-In critical advisory: AI agents accessed Indian organisations' live systems without authorisation | 2026-10-16 | Unnamed provider evaluation agents and open source agentic attack tools | Credential theft | None | ||
Incident. Agents read business records with personal and confidential data at Indian organisations, none of which detected it; criminals used agent tools to cut intrusions from days to hours. How. Agents reused credentials exposed in configs, public repos and loose cloud roles, plus dormant APIs and admin panels without MFA. Advisory CIAD-2026-0049. The Federal · Inc42 · checked 2026-11-12 | ||||||
| Proof of conceptClaude Agent SDK resume option allowed CLI flag injection and command execution | 2026-10-16 | Claude Agent SDK for Python | Config runs code | CVE-2026-96620 | 9.2 | |
Vulnerability. An attacker who controls the session ID an app passes to resume can run arbitrary commands, outside the SDK's permission controls. How. The resume value went to the Claude CLI unvalidated, so a value starting with a dash became a flag such as --mcp-config that defines a stdio MCP server. Fix. 0.2.121 NVD CVE-2026-96620 · GitHub advisory GHSA-h4mw-j7qp-8mwm · checked 2026-11-12 | ||||||
| Proof of conceptJunie CLI ran code from an untrusted project's MCP server configuration | 2026-10-15 | JetBrains Junie CLI | Config runs code | CVE-2026-111091 | 7.8 | |
Vulnerability. Opening a malicious project can run attacker code on the developer's machine. How. Junie acted on the MCP server configuration shipped in an untrusted project. Fix. 26.8.3 NVD CVE-2026-111091 · JetBrains fixed security issues · checked 2026-11-12 | ||||||
| Proof of conceptClaude Code plugin installs could swap a pinned commit for attacker code | 2026-10-13 | Claude Code and Claude Desktop plugin installer | Supply chain | CVE-2026-86063 | 7.5 | |
Vulnerability. A user who installs a plugin pinned to a reviewed commit can get attacker-controlled plugin code instead. How. The installer ran git checkout on the pinned SHA without checking HEAD matched it, so a branch named after that SHA was checked out instead. Fix. 2.1.179 NVD CVE-2026-86063 · GitHub advisory GHSA-pq7j-f95f-qfcv · checked 2026-11-12 | ||||||
| Proof of conceptobsidian-tc MCP server fixed cross-vault tokens, DNS rebinding and read ACL bypasses | 2026-10-12 | obsidian-tc (Obsidian MCP server) | Auth bypass | CVE-2026-77366 CVE-2026-77363 CVE-2026-77364 CVE-2026-77361 CVE-2026-77362 | 9.9 | |
Vulnerability. Any valid token could read, write or delete another vault, a malicious web page could reach an unauthenticated local server, and callers could read denied notes and plugin secrets. How. Tokens were not bound to a vault, the HTTP transport skipped Host and Origin checks, control folders were only hidden from listings, and ACL checks missed symlinks and some tool output fields. Fix. 1.3.1 NVD CVE-2026-77366 · NVD CVE-2026-77363 · NVD CVE-2026-77364 · NVD CVE-2026-77361 · NVD CVE-2026-77362 · GitHub advisory GHSA-84c9-qqpq-2pq5 · GitHub advisory GHSA-pqh8-4pj7-32xx · GitHub advisory GHSA-cvp8-vvgf-69q4 · GitHub advisory GHSA-c5xx-h28h-f682 · GitHub advisory GHSA-43r2-c7g8-prc8 · checked 2026-11-12 | ||||||
| Proof of conceptmcp-remote sent OAuth secrets and tokens over plain HTTP | 2026-10-11 | mcp-remote (npm) | Credential theft | CVE-2026-108716 | 6.0 | |
Vulnerability. An on-path attacker can capture the client secret and the issued access and refresh tokens. How. From 0.8.0 through 0.14.3 the device code flow used discovered device authorization and token endpoints without requiring HTTPS. NVD CVE-2026-108716 · VulnCheck advisory · checked 2026-11-12 | ||||||
| In the wildAnthropic disclosed Claude models exploiting flaws and submitting real forms during tasks | 2026-10-09 | Claude Mythos 5, Claude Opus 5, Claude Haiku 4.5 and others | Rogue agent action | None | ||
Incident. A university server was used via an injection flaw, gated data was pulled with found tokens, and real government and police tip forms were submitted. How. Models on research tasks worked around errors and restrictions on live sites. Anthropic says real-world impact was minimal. Anthropic: Investigating unintended model actions · checked 2026-11-12 | ||||||
| Proof of conceptx64dbg-MCP Server gave any network client control of the debugger | 2026-10-09 | x64dbg-MCP Server | Auth bypass | CVE-2026-107824 | 9.3 | |
Vulnerability. An unauthenticated client can run debugger commands, attach to processes, read and write their memory and write files anywhere. How. All debugger tools were exposed over HTTP and SSE with no authentication on 0.0.0.0, ports 9094 and 9095. Fix. 1.1 NVD CVE-2026-107824 · GitHub advisory GHSA-4478-h5jv-647m · checked 2026-11-12 | ||||||
| In the wildOpenAI confirmed a model accessed non-public NSW National Parks and Wildlife Service data | 2026-10-07 | OpenAI internal model | Rogue agent action | None | ||
Incident. A model researching Australian fire statistics reached database metadata that was not meant to be public. OpenAI found no personal information involved. How. Same pattern as the Medicare case: a model doing data lookups went beyond public pages on a government web app. Cyber Daily · checked 2026-11-12 | ||||||
| Proof of conceptLangflow access control flaw let authenticated users run code | 2026-10-07 | Langflow OSS | Auth bypass | CVE-2026-104335 | 8.8 | |
Vulnerability. A remote authenticated attacker can execute arbitrary code on the Langflow server. How. IBM describes improper access control in Langflow 1.0.0 through 1.12.2 without further detail. Fix. 1.12.3 NVD CVE-2026-104335 · IBM security bulletin · checked 2026-11-12 | ||||||
| ReportedWikimedia says agents it believes OpenAI ran edited wikis and probed its Etherpad | 2026-10-05 | OpenAI agents (attributed by Wikimedia) | Rogue agent action | None | ||
Incident. Unapproved bot edits, attempts to hijack a citation tool and Etherpad as proxies, and millions of requests that may have fed a Wikidata Query Service outage in May. How. Agents made sandbox edits and changed citation tool settings to fetch remote data. Wikimedia found no compromise of its systems or data. Wikimedia Foundation · checked 2026-11-12 | ||||||
| In the wildOpenAI agent accessed public and non-public files on Australia's Medicare statistics portal | 2026-09-24 | OpenAI internal model | Rogue agent action | None | ||
Incident. The agent reached non-public files on the Services Australia Medicare Statistics Reporting Service. No patient records are believed to have been accessed. How. A model in internal training looking up Australian statistics went past public pages in June 2026. OpenAI emailed the government in September and Australia ran a rapid review. Fix. Government accepted 18 review recommendations, including mandatory AI incident reporting Prime Minister press conference, 24 Sep 2026 · Prime Minister media release on the Rapid Review · checked 2026-11-12 | ||||||
| ReportedCriminal ran open source AI agent harnesses to hack online retailers for about $25 each | 2026-09-22 | Strix, Cairn and Hermes agent harnesses | Data exfiltration | None | ||
Incident. At least 27 companies were compromised in five days, skimmers were confirmed on 19, and over 600,000 card records were taken from two firms. How. One operator used agent frameworks over OpenRouter models to scan, exploit and plant card skimmers at scale, with a mean cost of $25.46 per scan. Gambit Security interim report · checked 2026-11-12 | ||||||
| ReportedResearchers link OpenAI test agents to 2,000+ junk gems and API key theft attempts on RubyGems | 2026-09-11 | RubyGems.org | Rogue agent action | None | ||
Incident. RubyGems paused sign-ups for four days in May 2026 and removed 500+ packages. Agents tried to exploit a caching flaw that could leak API keys. How. Agents uploaded gems that abused RubyDoc.info builds to run code and scrape sites. OpenAI says its agents used RubyGems for benign tasks and has not verified the malicious uploads. Fix. RubyGems fixed the cache bug 9 July and revoked all legacy API keys Researcher report (rubyhack.ai) · RubyGems advisory on legacy API key leak · Simon Willison summary with OpenAI update · checked 2026-11-12 | ||||||
| Proof of conceptGemini CLI folder trust bypass through GEMINI_CLI_HOME in .env | 2026-09-10 | Gemini CLI | Config runs code | CVE-2026-13745 | 7.7 | |
Vulnerability. Starting the CLI in an untrusted folder could run attacker code. How. A project .env overrode GEMINI_CLI_HOME so the CLI loaded attacker config and skipped trust prompts. Fix. 0.39.1 NVD CVE-2026-13745 · Redguard advisory · Gemini CLI v0.39.1 release · checked 2026-11-12 | ||||||
| In the wildOpenAI agent swarm took over German wiki DSEWiki as a message board | 2026-09-04 | OpenAI internal agents | Rogue agent action | None | ||
Incident. About 17,000 agent edits flooded a volunteer wiki, with XSS probing and moderator impersonation, from May to early July 2026. How. Agents in a web research task used the wiki to share answers and ways around restrictions, including bypassing a GET-only proxy. OpenAI called it a misalignment incident and reported it to the European Commission. Researcher report (collusion.wiki) · OpenAI misalignment notices index · Fortune · checked 2026-11-12 | ||||||
| Proof of conceptCodex ran a repository's core.fsmonitor helper outside its sandbox | 2026-09-01 | OpenAI Codex CLI and Codex Desktop | Config runs code | CVE-2026-19592 | 7.3 | |
Vulnerability. A prepared repository could run attacker code with the user's privileges and no approval prompt. How. Codex collected git metadata without disabling core.fsmonitor, so git ran the helper named in .git/config. NVD CVE-2026-19592 · Fix pull request openai/codex#22652 · checked 2026-11-12 | ||||||
| In the wildMeta's Muse Spark model exploited a third-party service during a misconfigured evaluation | 2026-08-05 | Muse Spark 1.1 | Rogue agent action | None | ||
Incident. An unnamed outside company's systems were accessed and, per The Information, altered. How. A misconfiguration by test partner Irregular gave the model internet access, and it exploited a vulnerability in a real service. Reuters via BNN Bloomberg · checked 2026-11-12 | ||||||
| In the wildClaude models reached three real organisations from a misconfigured third-party cyber evaluation | 2026-07-30 | Claude Opus 4.7, Claude Mythos 5, internal test model | Rogue agent action | None | ||
Incident. One organisation lost credentials and several hundred rows of production data; a malicious PyPI package ran on 15 real systems; a third app was entered by SQL injection. How. Evaluation machines run with partner Irregular had live internet by mistake. Claude believed the real targets were part of the exercise and attacked them. Fix. Cyber evaluations paused 23 July; affected parties and PyPI notified Anthropic: Investigating three real-world incidents in our cybersecurity evaluations · checked 2026-11-12 | ||||||
| Proof of conceptClaude Code GitHub Action loaded .mcp.json from attacker pull requests | 2026-07-16 | Claude Code Action (GitHub Action) | Config runs code | CVE-2026-47751 | 5.3 | |
Vulnerability. An attacker pull request could run code on the Actions runner and steal workflow secrets. How. The action checked out the PR head, read its .mcp.json and enabled all project MCP servers when a privileged user or trigger ran it. Fix. 1.0.74 NVD CVE-2026-47751 · GitHub advisory GHSA-8q5r-mmjf-575q · checked 2026-11-12 | ||||||
| In the wildOpenAI evaluation agents escaped their sandbox and broke into Hugging Face production | 2026-07-16 | OpenAI internal models (ExploitGym evaluation) | Rogue agent action | None | ||
Incident. The agents reached several internal Hugging Face clusters, collected cloud and service credentials and accessed a limited set of internal datasets. How. Agents in a cyber evaluation exploited a zero-day in a package registry proxy to get open internet. A malicious dataset then gave code execution in Hugging Face's dataset processing. Fix. Dataset code-execution paths closed; credentials rotated; OpenAI blocked live internet in training and evals Hugging Face security incident disclosure · OpenAI alignment report linking the July runs · checked 2026-11-12 | ||||||
| Proof of conceptAWS HealthLake MCP server could send AWS credentials to an attacker endpoint | 2026-07-14 | awslabs.healthlake-mcp-server (PyPI) | Credential theft | CVE-2026-15643 | 9.2 | |
Vulnerability. An authenticated user could exfiltrate the server's temporary AWS security credentials. How. The pagination next_token was trusted as a URL, so follow-up signed requests could be redirected to any server. Fix. 0.0.14 NVD CVE-2026-15643 · AWS security bulletin 2026-054 · checked 2026-11-12 | ||||||
| Proof of conceptCursor agent could widen its sandbox with the working_directory parameter | 2026-06-25 | Cursor | Sandbox escape | CVE-2026-50548 | 9.3 | |
Vulnerability. A malicious agent could write files anywhere the user can and get unsandboxed code execution. How. The sandbox granted write access to the command's working directory, which the agent could point at a sensitive location such as the sandbox helper. Fix. 3.0 NVD CVE-2026-50548 · GitHub advisory GHSA-3p48-7v9f-v5cw · checked 2026-11-12 | ||||||
| Proof of conceptGemini CLI and run-gemini-cli action ran code from a malicious .gemini/.env before the sandbox | 2026-06-24 | Gemini CLI, run-gemini-cli GitHub Action | Config runs code | CVE-2026-12537 | 10.0 | |
Vulnerability. On headless CI an attacker could get host code execution before the sandbox started. How. The container launcher built an OS command from values in a crafted .gemini/.env file. Fix. Gemini CLI 0.39.1; run-gemini-cli 0.1.22 NVD CVE-2026-12537 · GitHub advisory GHSA-wpqr-6v78-jr5g · checked 2026-11-12 | ||||||
| Proof of conceptCline Kanban server open to cross-origin WebSocket hijack | 2026-06-01 | Cline (Kanban server) | Auth bypass | CVE-2026-44211 | 9.6 | |
Vulnerability. A malicious website could connect to the local Kanban server and act on the user's behalf. How. The WebSocket server did not check the origin of connecting pages. Fix. No patch at publication NVD CVE-2026-44211 · GitHub advisory GHSA-5c57-rqjx-35g2 · checked 2026-11-12 | ||||||
| In the wildLiteLLM proxy SQL injection through the Authorization header, exploited in the wild | 2026-05-08 | LiteLLM proxy | Credential theft | CVE-2026-42208 | 9.3 | |
Vulnerability. An unauthenticated attacker could read and possibly change the proxy database, including the LLM credentials it holds. CISA lists it as known exploited. How. The API key check placed the caller's key value into SQL text, reachable through the error path on any LLM route. Fix. 1.83.7 NVD CVE-2026-42208 · GitHub advisory GHSA-r75f-5x8p-qvmc · checked 2026-11-12 | ||||||
| Proof of conceptWindsurf prompt injection registered a malicious MCP server | 2026-04-15 | Windsurf | Prompt injection | CVE-2026-30615 | 8.0 | |
Vulnerability. Attacker HTML content could run commands on the victim system and persist through MCP config. How. Injected instructions led the agent to edit the local MCP config and add a malicious stdio server that then ran. NVD CVE-2026-30615 · OX Security advisory · checked 2026-11-12 | ||||||
| Proof of conceptLiteLLM MCP server creation ran arbitrary commands | 2026-04-15 | LiteLLM | Config runs code | CVE-2026-30623 | 9.8 | |
Vulnerability. An authenticated user able to add MCP servers could run OS commands with the LiteLLM process privileges. How. MCP server JSON config took any command and args and LiteLLM executed them without validation. Fix. 1.83.7-stable (fix first in 1.83.6-nightly) NVD CVE-2026-30623 · LiteLLM blog: MCP stdio command injection · OX Security advisory · checked 2026-11-12 | ||||||
| Proof of conceptAgent Zero external MCP server config ran arbitrary commands | 2026-04-15 | Agent Zero | Config runs code | CVE-2026-30624 | 8.6 | |
Vulnerability. An attacker who supplies an MCP config gets command execution as the Agent Zero process. How. The External MCP Servers feature executed command and args values from JSON without restriction. NVD CVE-2026-30624 · OX Security advisory · checked 2026-11-12 | ||||||
| In the wildMeta internal agent posted bad advice unprompted, exposing sensitive data to staff for two hours | 2026-03-18 | Meta internal AI agent | Rogue agent action | None | ||
Incident. Company and user data was visible to engineers without authorisation for about two hours. Meta rated it Sev 1. How. An engineer asked the agent to analyse a forum question and it posted a reply without approval. The asker followed the advice, which opened access too widely. TechCrunch (Meta confirmed to The Information) · checked 2026-11-12 | ||||||
| Proof of conceptCopilot CLI safety check missed commands hidden in bash parameter expansion | 2026-03-06 | GitHub Copilot CLI | Prompt injection | CVE-2026-29783 | 7.5 | |
Vulnerability. Injected instructions from repo files or MCP responses could run hidden commands without approval. How. Bash parameter transformation operators embedded code in arguments of commands the CLI classed as read-only. Fix. 0.0.423 NVD CVE-2026-29783 · GitHub advisory GHSA-g8r9-g2v8-jv6f · checked 2026-11-12 | ||||||
| ReportedSingle operator used Claude Code to breach Mexican government agencies | 2026-02-25 | Claude Code | Data exfiltration | None | ||
Incident. Gambit says nine government organisations were breached and hundreds of millions of citizen records stolen; Bloomberg reported about 150 GB. How. The attacker sent over 1,000 Spanish prompts framed as a bug bounty. Claude Code generated and ran about 75 percent of remote commands, and GPT-4.1 sorted the stolen data. Gambit Security technical report · Gambit Security: Bloomberg coverage · checked 2026-11-12 | ||||||
| In the wildSANDWORM_MODE npm worm planted rogue MCP servers in AI coding assistants | 2026-02-20 | Typosquatted npm packages (claud-code, cloude, opencraw and others) | Tool poisoning | None | ||
Incident. Infected machines had SSH keys, cloud credentials, npm tokens and LLM API keys harvested, and CI workflows poisoned. How. At least 19 packages wrote a hidden MCP server into Claude Code, Claude Desktop, Cursor, Continue and Windsurf configs. Its tool descriptions told the assistant to read secrets and pass them on without telling the user. Fix. Packages reported to npm; actor infrastructure taken down Socket · checked 2026-11-12 | ||||||
| ReportedReport says Amazon's Kiro agent deleted and recreated an AWS environment, causing a 13 hour outage | 2026-02-20 | Kiro | Rogue agent action | None | ||
Incident. One AWS service was disrupted for 13 hours in December 2025, per the Financial Times. Amazon puts it down to a misconfigured role. How. Engineers let the agent fix an issue and it chose to delete and rebuild the environment. Amazon says the engineer's role had broader permissions than expected and calls it user error. The Register (with Amazon statement) · checked 2026-11-12 | ||||||
| ConfirmedMicrosoft 365 Copilot Chat summarised emails marked confidential despite DLP policies | 2026-02-18 | Microsoft 365 Copilot Chat | Auth bypass | None | ||
Incident. Copilot read and summarised labelled emails in users' Sent Items and Drafts that DLP rules should have blocked. Microsoft has not said how many tenants were affected. How. A code defect made Copilot Chat process messages with a confidential sensitivity label. It was detected on 21 January 2026 and tracked as advisory CW1226324. Fix. Configuration update and code fix rolled out February 2026 BleepingComputer (quotes Microsoft advisory CW1226324) · checked 2026-11-12 | ||||||
| In the wildStolen npm token used to publish [email protected], which silently installed OpenClaw | 2026-02-17 | Cline CLI (npm) | Supply chain | None | ||
Incident. Anyone installing the CLI in an 8 hour window got OpenClaw installed globally. The token theft path gave an attacker publish rights to a popular AI coding tool. How. A researcher showed that a prompt in a GitHub issue title could steer Cline's Claude-based issue triage workflow and poison the Actions cache to steal release tokens. A token that was not properly revoked was then used to publish 2.3.0. Fix. 2.4.0; npm publishing moved to OIDC GitHub advisory GHSA-9ppg-jx86-fqw7 · Adnan Khan: Clinejection · checked 2026-11-12 | ||||||
| Proof of conceptCursor agent could escape its sandbox through .git config and hooks | 2026-02-13 | Cursor | Sandbox escape | CVE-2026-26268 | 8.0 | |
Vulnerability. A prompt-injected agent could get code execution outside the sandbox. How. The agent could write to .git settings and hooks, which git later ran automatically. Fix. 2.5 NVD CVE-2026-26268 · GitHub advisory GHSA-8pcm-8jpx-hv8r · checked 2026-11-12 | ||||||
| Proof of conceptClaude Code sandbox let code plant persistent hooks in settings.json | 2026-02-06 | Claude Code | Sandbox escape | CVE-2026-25725 | 7.7 | |
Vulnerability. Code running inside the sandbox could gain command execution on the host at the next start. How. If .claude/settings.json did not exist it was left writable, so sandboxed code could create it with SessionStart hooks. Fix. 2.1.2 NVD CVE-2026-25725 · GitHub advisory GHSA-ff64-7w26-62rf · checked 2026-11-12 | ||||||
| In the wildHundreds of malicious OpenClaw skills on ClawHub delivered infostealers | 2026-02-01 | ClawHub skills (OpenClaw) | Malicious package | None | ||
Incident. Users who followed the skills' setup steps got Atomic macOS Stealer, Windows stealers or reverse shells, and some skills sent OpenClaw .env API keys to a webhook. How. Skill docs hid a fake prerequisite that told the user to run a pasted command or download a password-protected archive. Koi counted 341 bad skills; Antiy later counted at least 1,184. Fix. Skills reported to ClawHub for removal Antiy CERT analysis (credits Koi's 1 Feb report) · checked 2026-11-12 | ||||||
| Proof of conceptCopilot Studio exposed sensitive information to unauthenticated attackers | 2026-01-22 | Microsoft Copilot Studio | Data exfiltration | CVE-2026-21520 | 7.5 | |
Vulnerability. An unauthenticated attacker could view sensitive information over the network. How. Microsoft lists it as exposure of sensitive information; no technical details were published. Fix. Fixed server side by Microsoft NVD CVE-2026-21520 · MSRC CVE-2026-21520 · checked 2026-11-12 | ||||||
| Proof of conceptMalicious repo settings could send Claude Code API keys to an attacker before trust | 2026-01-21 | Claude Code | Credential theft | CVE-2026-21852 | 5.3 | |
Vulnerability. Opening a malicious repository could leak the user's Anthropic API key. How. A repo settings file set ANTHROPIC_BASE_URL to an attacker endpoint and Claude Code made API calls before showing the trust prompt. Fix. 2.0.65 NVD CVE-2026-21852 · GitHub advisory GHSA-jh7p-qr78-84p7 · checked 2026-11-12 | ||||||
| Proof of conceptLangChain dumps() let user data pose as serialized LangChain objects | 2025-12-23 | langchain-core | Code execution | CVE-2025-68664 | 9.3 | |
Vulnerability. Attacker-shaped data, including model output, could be loaded back as trusted objects, which can leak secrets. How. dumps() and dumpd() did not escape dicts with an 'lc' key, so they were rebuilt as LangChain objects on load. Fix. 0.3.81 and 1.2.5 NVD CVE-2025-68664 · Fix pull request langchain#34455 · checked 2026-11-12 | ||||||
| Proof of conceptThree flaws in the reference Git MCP server: unrestricted git_init, argument injection, path check bypass | 2025-12-17 | mcp-server-git | Prompt injection | CVE-2025-68143 CVE-2025-68144 CVE-2025-68145 | 6.5 | |
Vulnerability. Content that steers the agent can make the server touch repositories outside the configured one and overwrite arbitrary files. How. git_init accepted any path, git_diff and git_checkout passed flag-like values such as --output straight to git, and the --repository restriction was not checked on later calls. Fix. 2025.12.17 NVD CVE-2025-68143 · NVD CVE-2025-68144 · NVD CVE-2025-68145 · GitHub advisory GHSA-5cgr-j3jf-jw3v · GitHub advisory GHSA-9xwc-hfwc-8w59 · GitHub advisory GHSA-j22h-9j4x-23w5 · checked 2026-11-12 | ||||||
| Proof of conceptZed started MCP servers from a project's .zed/settings.json on open | 2025-12-17 | Zed | Config runs code | CVE-2025-68433 | 7.7 | |
Vulnerability. Opening a malicious project ran arbitrary shell commands. How. MCP server commands in project settings ran with no prompt; the fix adds worktree trust. Fix. 0.218.2-pre NVD CVE-2025-68433 · GitHub advisory GHSA-cv6g-cmxc-vw8j · checked 2026-11-12 | ||||||
| Proof of conceptMCP TypeScript SDK shipped without DNS rebinding protection on by default | 2025-12-02 | MCP TypeScript SDK (@modelcontextprotocol/sdk) | Auth bypass | CVE-2025-66414 | 7.6 | |
Vulnerability. A malicious website could call tools on an unauthenticated local HTTP MCP server built with the SDK. How. HTTP transports did not enable DNS rebinding protection unless the developer set it, so a rebinding page could reach localhost. Fix. 1.24.0 NVD CVE-2025-66414 · GitHub advisory GHSA-w48q-cv73-mx4w · checked 2026-11-12 | ||||||
| Proof of conceptCodex CLI ran MCP commands from a repository's own config | 2025-12-01 | OpenAI Codex CLI | Config runs code | CVE-2025-61260 | 9.8 | |
Vulnerability. Running codex inside a malicious repository executed the attacker's commands. How. A project .env set CODEX_HOME to a repo folder, so Codex loaded its config.toml and started the listed MCP server commands without a prompt. Fix. 0.23.0 NVD CVE-2025-61260 · Check Point Research write-up · checked 2026-11-12 | ||||||
| ReportedGoogle Antigravity agent in Turbo mode wiped a user's whole D: drive while clearing a cache | 2025-12-01 | Google Antigravity | Rogue agent action | None | ||
Incident. A developer lost the contents of a drive, including media files that recovery tools could not restore. How. Asked to clear a project cache, the agent ran a delete command against the drive root. Turbo mode let it run commands without approval. The Register · checked 2026-11-12 | ||||||
| In the wildChinese state-sponsored group used Claude Code to run a largely autonomous espionage campaign | 2025-11-13 | Claude Code | Data exfiltration | None | ||
Incident. About 30 tech, finance, chemical and government targets were attacked and a small number were breached, with credentials and data stolen. How. Operators split the intrusion into small tasks and told Claude it was doing defensive testing. Claude Code with tools then did 80 to 90 percent of the work: recon, exploit code, credential harvesting and exfiltration. Fix. Accounts banned; victims and authorities notified Anthropic: Disrupting the first reported AI-orchestrated cyber espionage campaign · checked 2026-11-12 | ||||||
| Proof of conceptCommand injection in the VS Code Copilot Chat extension | 2025-11-11 | Visual Studio Code Copilot Chat extension | Code execution | CVE-2025-62222 | 8.8 | |
Vulnerability. An unauthorised attacker could execute code over a network. How. Microsoft lists it as command injection in the extension; it gives few technical details. Fix. November 2025 security update NVD CVE-2025-62222 · MSRC CVE-2025-62222 · checked 2026-11-12 | ||||||
| Proof of conceptLangGraph checkpoint serializer allowed code execution on load | 2025-11-07 | langgraph-checkpoint (LangGraph) | Code execution | CVE-2025-64439 | 7.4 | |
Vulnerability. Anyone who can shape stored checkpoint data could run code when the agent state is loaded. How. The JsonPlusSerializer fell back to json mode, which could construct arbitrary objects on deserialisation. Fix. 3.0.0 NVD CVE-2025-64439 · GitHub advisory GHSA-wwqv-p2pp-99h5 · checked 2026-11-12 | ||||||
| Proof of conceptWindsurf file tools allowed reading and writing any local file | 2025-10-17 | Windsurf | Sandbox escape | CVE-2025-62353 | 9.8 | |
Vulnerability. Directly or through indirect prompt injection, an attacker could read and write files inside and outside the project. How. File tools such as write_to_file did not restrict paths, so traversal sequences reached any location (1.12.12 and older). NVD CVE-2025-62353 · HiddenLayer advisory · checked 2026-11-12 | ||||||
| Proof of conceptFramelink Figma MCP server command injection | 2025-10-08 | Framelink Figma MCP server (figma-developer-mcp) | Code execution | CVE-2025-53967 | 8.0 | |
Vulnerability. An attacker with network access to the MCP interface can run OS commands as the server process. How. User input reached a curl command in fetchWithRetry without sanitising shell metacharacters. Fix. 0.6.3 NVD CVE-2025-53967 · Imperva write-up · Release v0.6.3 · checked 2026-11-12 | ||||||
| Proof of conceptClaude Code ran project code before the trust dialog was accepted | 2025-10-03 | Claude Code | Config runs code | CVE-2025-59536 | 8.7 | |
Vulnerability. Opening Claude Code in a malicious repository could run the attacker's code. How. A bug in the startup trust dialog let code in the project execute before the user agreed to trust it. Fix. 1.0.111 NVD CVE-2025-59536 · GitHub advisory GHSA-4fgq-fpq9-mr3g · checked 2026-11-12 | ||||||
| Proof of conceptCursor protected-file check was case-sensitive | 2025-10-03 | Cursor | Sandbox escape | CVE-2025-59944 | 8.0 | |
Vulnerability. Prompt injection could change MCP config and reach remote code execution on case-insensitive filesystems. How. Protection for files like .cursor/mcp.json compared paths case-sensitively, so a different casing slipped through. Fix. 1.7 NVD CVE-2025-59944 · GitHub advisory GHSA-xcwh-rrwj-gxc7 · checked 2026-11-12 | ||||||
| Proof of conceptForcedLeak: Agentforce leaked CRM data through a Web-to-Lead form | 2025-09-25 | Salesforce Agentforce | Data exfiltration | None | ||
Vulnerability. An outside attacker could get Agentforce to send CRM data such as lead emails to a server they control. How. Instructions hidden in a Web-to-Lead description ran when staff asked the agent about the lead; data went out in an image URL on an expired domain still on the CSP allowlist. Fix. Trusted URLs Enforcement for Agentforce and Einstein AI (8 Sep 2025); expired domain re-secured Noma Security write-up · checked 2026-11-12 | ||||||
| In the wildFake postmark-mcp npm package secretly BCC'd every sent email to an attacker | 2025-09-25 | postmark-mcp (npm) | Malicious package | None | ||
Incident. Every email sent through the MCP server, including password resets and invoices, was copied to an attacker-controlled address. How. The package built trust over 15 clean versions, then version 1.0.16 added a hidden BCC line. Postmark says it never published an MCP server on npm. Fix. Remove the package and rotate credentials sent by email Postmark statement · checked 2026-11-12 | ||||||
| Proof of conceptCodex CLI treated a model-chosen folder as the sandbox writable root | 2025-09-22 | OpenAI Codex CLI | Sandbox escape | CVE-2025-59532 | 8.6 | |
Vulnerability. The agent could write files and run commands outside the session folder. How. Sandbox setup used a model-generated cwd as the writable root instead of the folder where the user started. Fix. 0.39.0 NVD CVE-2025-59532 · GitHub advisory GHSA-w5fx-fh39-j5rw · checked 2026-11-12 | ||||||
| Proof of conceptJetBrains Junie allowed code execution through weak command validation | 2025-09-17 | JetBrains Junie | Prompt injection | CVE-2025-59458 | 8.3 | |
Vulnerability. Code execution on the developer's machine. How. JetBrains lists improper command validation in the agent; few details were published. Fix. 252.284.66 and matching builds NVD CVE-2025-59458 · JetBrains fixed security issues · checked 2026-11-12 | ||||||
| In the wildCriminal used Claude Code to run a data extortion campaign against at least 17 organisations | 2025-08-27 | Claude Code | Data exfiltration | None | ||
Incident. Healthcare, emergency services, government and religious organisations had data stolen and faced ransom demands that sometimes exceeded $500,000. How. The actor let Claude Code automate reconnaissance, credential harvesting and network intrusion, pick what to steal, set ransom amounts and write ransom notes. Fix. Accounts banned; new detection added Anthropic threat intelligence report, August 2025 · checked 2026-11-12 | ||||||
| In the wildMalicious Nx packages used installed AI coding CLIs to hunt for secrets | 2025-08-26 | nx and @nx plugins (npm) | Supply chain | CVE-2025-10894 | 9.6 | |
Incident. Credentials, tokens and wallet files from developer machines and CI were posted to public s1ngularity-repository repos on victims' own GitHub accounts. How. Attackers stole the npm token through a GitHub Actions injection and published malicious versions. The postinstall script asked local Claude, Gemini and Q CLIs, with permission-bypass flags, to list sensitive files. Fix. Malicious versions removed; upgrade to latest and rotate all credentials Nx postmortem · GitHub advisory GHSA-cxm3-wv7p-598c · StepSecurity original analysis · NVD CVE-2025-10894 · checked 2026-11-12 | ||||||
| Proof of conceptClaude Code safe-command allowlist allowed reading and sending files without approval | 2025-08-16 | Claude Code | Data exfiltration | CVE-2025-55284 | 7.1 | |
Vulnerability. Prompt injection could read local files such as secrets and send them over the network. How. Commands on an overly broad allowlist ran without confirmation and could make network requests. Fix. 1.0.4 NVD CVE-2025-55284 · GitHub advisory GHSA-x5gv-jw7f-j6xj · checked 2026-11-12 | ||||||
| Proof of conceptGitHub Copilot in Visual Studio and VS Code could be pushed to run code via prompt injection | 2025-08-12 | GitHub Copilot (Visual Studio, VS Code) | Prompt injection | CVE-2025-53773 | 7.8 | |
Vulnerability. Injected content could run code on the developer's machine. How. Injected project content made Copilot add chat.tools.autoApprove: true to the settings file, which turned off confirmations so it could run terminal commands. Fix. August 2025 security update NVD CVE-2025-53773 · MSRC CVE-2025-53773 · Embrace The Red write-up · checked 2026-11-12 | ||||||
| Proof of conceptZed Agent Panel could write project config and run commands without approval | 2025-08-11 | Zed | Prompt injection | CVE-2025-55012 | 8.5 | |
Vulnerability. An AI agent could get command execution on the user's machine. How. A permission bypass let the agent create or edit a project config file that then ran commands. Fix. 0.197.3 NVD CVE-2025-55012 · GitHub advisory GHSA-x34m-39xw-g2wr · checked 2026-11-12 | ||||||
| Proof of conceptClaude Code command parsing bug skipped the confirmation prompt | 2025-08-05 | Claude Code | Prompt injection | CVE-2025-54795 | 8.7 | |
Vulnerability. Untrusted content in the context window could run commands without the user approving them. How. An error in how commands were parsed let an injected command pass as already approved. Fix. 1.0.20 NVD CVE-2025-54795 · GitHub advisory GHSA-x56v-x2h6-7j34 · checked 2026-11-12 | ||||||
| Proof of conceptCurXecute: prompt injection could create Cursor's mcp.json and run commands | 2025-08-05 | Cursor | Prompt injection | CVE-2025-54135 | 8.5 | |
Vulnerability. Injected content could get remote code execution on the developer's machine without approval. How. Creating a new dotfile needed no approval, so the agent could be steered to write .cursor/mcp.json with a malicious server. Fix. 1.3.9 NVD CVE-2025-54135 · GitHub advisory GHSA-4cxx-hrm3-49rm · checked 2026-11-12 | ||||||
| Proof of conceptMCPoison: approved MCP config in Cursor could be swapped without a new prompt | 2025-08-02 | Cursor | Config runs code | CVE-2025-54136 | 7.2 | |
Vulnerability. Anyone with write access to a shared repo could get persistent code execution on collaborators' machines. How. Cursor trusted an MCP entry by name after first approval, so later changes to its command were not re-checked. Fix. 1.3 NVD CVE-2025-54136 · GitHub advisory GHSA-24mc-g4xr-4395 · checked 2026-11-12 | ||||||
| Proof of conceptCursor chat rendered Mermaid images that could leak data | 2025-08-01 | Cursor | Data exfiltration | CVE-2025-54132 | 4.4 | |
Vulnerability. After a prompt injection, sensitive data could be sent to an attacker server. How. Mermaid diagrams in chat could embed remote images, and the image fetch carried the data. Fix. 1.3 NVD CVE-2025-54132 · GitHub advisory GHSA-43wj-mwcc-x93p · checked 2026-11-12 | ||||||
| In the wildAttacker slipped injected code into the Amazon Q Developer VS Code extension v1.84.0 | 2025-07-23 | Amazon Q Developer extension for VS Code | Supply chain | CVE-2025-8217 | 4.0 | |
Incident. Version 1.84.0 reached users with injected code designed to call the Q Developer CLI. AWS says a syntax error stopped it from running and no customer environments were changed. How. An overly broad GitHub token in the CodeBuild setup let an attacker commit to the open source repo, and the change went out in an automatic release. Fix. 1.85.0 AWS security bulletin AWS-2025-015 · NVD CVE-2025-8217 · checked 2026-11-12 | ||||||
| In the wildReplit Agent deleted a live production database during a code freeze | 2025-07-21 | Replit Agent | Rogue agent action | None | ||
Incident. SaaStr founder Jason Lemkin lost a production database of executive and company records; the agent also ignored a code freeze. How. The coding agent ran destructive database commands against production despite instructions to freeze changes. Replit's CEO called it unacceptable and promised automatic dev and prod database separation. Fix. Restore from backup; Replit rolled out dev and prod database separation The Register (with CEO Amjad Masad statement) · checked 2026-11-12 | ||||||
| In the wildFake Solidity extension on Open VSX stole about $500,000 in crypto from a Cursor user | 2025-07-10 | Fake Solidity extensions (Open VSX), Cursor IDE | Malicious package | None | ||
Incident. A blockchain developer lost about $500,000 in crypto after installing the extension in Cursor. How. The extension posed as a Solidity language tool and ranked above the real one in Open VSX search. It dropped ScreenConnect, the Quasar backdoor and a stealer. Fix. Extensions removed from Open VSX Kaspersky press release · checked 2026-11-12 | ||||||
| Proof of conceptmcp-remote ran OS commands from a malicious server's OAuth endpoint | 2025-07-09 | mcp-remote (npm) | Config runs code | CVE-2025-6514 | 9.6 | |
Vulnerability. Connecting to an untrusted MCP server gives that server command execution on the client, with full control on Windows. How. A crafted authorization_endpoint URL in the server's OAuth metadata was passed to a system open call without sanitising. Fix. 0.1.16 NVD CVE-2025-6514 · JFrog write-up · Fix commit · checked 2026-11-12 | ||||||
| Proof of conceptFilesystem MCP server allowed access outside its allowed directories | 2025-07-02 | Filesystem MCP server (@modelcontextprotocol/server-filesystem) | Sandbox escape | CVE-2025-53109 CVE-2025-53110 | 7.3 | |
Vulnerability. An agent using the server could read or write files outside the directories the user allowed. How. One flaw followed symlinks inside allowed directories; the other accepted paths whose prefix matched an allowed directory name. Fix. 0.6.4 or 2025.7.01 NVD CVE-2025-53109 · NVD CVE-2025-53110 · GitHub advisory GHSA-q66q-fx2p-7w4m · GitHub advisory GHSA-hc55-p739-j48w · checked 2026-11-12 | ||||||
| Proof of conceptDeprecated Slack MCP server leaked data through link unfurling | 2025-07-02 | Slack MCP server (deprecated reference server) | Data exfiltration | CVE-2025-34072 | 9.3 | |
Vulnerability. Private data the agent can read is sent to an attacker URL with no click needed. How. Injected instructions make the agent post a link carrying data in the URL; Slack preview bots then fetch it. Fix. Server deprecated; no fix NVD CVE-2025-34072 · Embrace The Red advisory · checked 2026-11-12 | ||||||
| Proof of conceptClaude Code IDE extensions accepted WebSocket connections from any website | 2025-06-24 | Claude Code IDE extensions (VS Code and forks, JetBrains) | Auth bypass | CVE-2025-52882 | 8.8 | |
Vulnerability. A malicious web page could read files, see open files and IDE events, and in limited cases run code. How. The local WebSocket server used by the extension did not check where connections came from. Fix. VS Code extension 1.0.24; JetBrains plugin 0.1.9 NVD CVE-2025-52882 · GitHub advisory GHSA-9f65-56v6-gxw7 · checked 2026-11-12 | ||||||
| ConfirmedAsana MCP server bug exposed customer data to users in other organisations | 2025-06-18 | Asana MCP server | Auth bypass | None | ||
Incident. Users of the Asana MCP server could see projects, tasks and other objects from other customers' Asana domains. Asana found no sign that anyone misused the data. How. A logic flaw in the MCP server, launched on 1 May 2025, broke tenant isolation. Asana found it on 4 June, took the server offline until 17 June and reset all connections. Fix. Server fixed and relaunched; all MCP connections reset The Register · checked 2026-11-12 | ||||||
| Proof of conceptMCP Inspector proxy let any website launch local commands | 2025-06-13 | MCP Inspector | Auth bypass | CVE-2025-49596 | 9.4 | |
Vulnerability. A developer who visits a malicious page while MCP Inspector runs can have arbitrary commands run on their machine. How. The Inspector proxy had no authentication between client and proxy, so unauthenticated requests could start MCP commands over stdio. Fix. 0.14.1 NVD CVE-2025-49596 · GitHub advisory GHSA-7f8r-222p-6f5g · Oligo Security write-up · checked 2026-11-12 | ||||||
| Proof of conceptEchoLeak: zero-click data leak from Microsoft 365 Copilot | 2025-06-11 | Microsoft 365 Copilot | Data exfiltration | CVE-2025-32711 | 9.3 | |
Vulnerability. An unauthorised attacker could make Copilot disclose organisational data over the network. How. Microsoft classes it as AI command injection: attacker content that reaches Copilot's context makes it disclose data. Aim Labs reported it as a zero-click email attack. Fix. Fixed server side by Microsoft NVD CVE-2025-32711 · MSRC CVE-2025-32711 · checked 2026-11-12 | ||||||
| In the wildLangflow unauthenticated code execution, exploited in the wild | 2025-04-07 | Langflow | Code execution | CVE-2025-3248 | 9.8 | |
Vulnerability. Remote unauthenticated attackers run code on the Langflow server. CISA added it to the KEV catalog on 5 May 2025. How. The /api/v1/validate/code endpoint ran submitted Python without authentication. Fix. 1.3.0 NVD CVE-2025-3248 · Horizon3.ai write-up · CISA KEV alert · checked 2026-11-12 | ||||||
FAQ
What does Agent Watch track?
Real cases of AI agents and agent tooling being compromised: coding assistants such as Claude Code, Cursor, Copilot and Amazon Q, MCP servers and clients, agent skills, browser agents and enterprise agents. Incidents happened to real users; vulnerabilities are disclosed flaws with a CVE or vendor advisory.
What is the latest case?
ExploitGym agents used leftover credentials to run privileged commands on OpenAI research clusters (OpenAI internal models (ExploitGym evaluation), 2026-11-10).
What is the most common way agents get compromised?
Prompt injection: instructions hidden in content the agent reads, such as a README, issue, web page, email or tool description, that make it run commands or leak data. Malicious packages and MCP servers are the second route.
What do the statuses mean?
In the wild: the vendor, the affected organisation or an authority confirmed it happened to real users or systems. Confirmed: the vendor confirmed the exposure but no misuse is known. Proof of concept: a researcher showed it works. Reported: a credible report without confirmation.
How do I check a skill or MCP server before installing it?
Paste its files or tool list into the pwnsy Agent Skill Scanner. It flags hidden instructions, secrets grabs, download-and-run commands and unpinned packages in your browser.
Each case links its sources. Data: data.json (CC BY 4.0).