Individual incidents; prevalence is unknown. The Amazon Q instruction failed to run and VirusTotal does not report installs; postmark-mcp is the clearest case of use against real users. The Nx case shows malware abusing installed agents, not a malicious agent package. The LiteLLM case is a compromised AI gateway library that agents and their operators depend on, not an agent extension. JFrog’s case is agent configuration hooks planted in hijacked GitHub repositories, not packages in a registry, and does not establish execution on developer machines.
Corroborated: Supported by at least two independent sources.
Evidence
The injected instruction was malformed and did not run.
Not an agent package: a compromised build-tool package whose install script tried to use locally installed AI coding CLIs.
An impersonating MCP server on npm; version 1.0.16 blind-copied sent emails to the publisher.
Hundreds of malicious skills among 3,016+ analyzed, 314 from one publisher; install counts not reported.
A compromised release of the LiteLLM AI gateway on PyPI and Docker registries was built to steal cloud, developer and AI coding tool credentials; Wiz confirms abuse of stolen cloud credentials.
Cursor traces the tampered extension builds to a credential exposed at Fireworks and unredacted credentials in historical training traces. It reports at most 5,000 potential recipients. That ceiling does not establish successful remote access on every host.
JFrog reports 42 hijacked GitHub repositories linked to an npm campaign carrying execution hooks for Claude, Gemini, Cursor and VS Code; the count does not establish how many developer machines ran them.
Key questions that rely on this finding
- Where are deployed AI agents actually being exploited?Exploitation often reaches deployed agents through connectors, credentials, tools and packages.
- How are attackers using AI agents in real operations?To run large parts of intrusions under human direction, including espionage, credential theft and extortion; reported autonomy and success vary.
Status history
- 2025-07-23ReportedA malicious prompt shipped in an Amazon Q extension release. · record
- 2025-08-26CorroboratedThe Nx compromise abused installed AI coding CLIs. · record
- 2026-09-25CorroboratedcorrectionThe Nx compromise was a malicious build-tool package that invoked installed AI CLIs, not a malicious agent extension, MCP server or skill. Corroboration rests on the malicious postmark-mcp server (found by Koi Security, disclosed by Postmark), independent of the Amazon Q incident. · record