What happened
Bitdefender found 22 GitHub repositories whose instruction files tell AI coding agents like Claude, Codex and Copilot to silently run a credential-stealing program. The instructions tell the agent not to read the code and not to ask the developer. The malware looks for the settings of 14 agents, from Cursor and Gemini to Amazon Q.
Why it matters
The malware grabs cloud and GitHub tokens, SSH keys and password-manager data. Using the stolen tokens, it tries to plant the same trap in up to 25 of the victim's own repos and republish infected packages to npm and PyPI, spreading the attack. Any company whose developers use AI coding agents is exposed.
What to do
Review the instruction files of repositories processed by AI agents and block silent execution of binaries without developer approval. Revoke cloud and GitHub tokens and rotate SSH keys on machines that ran agents over unknown repositories.
For the technical team
- Affected
- Repositórios GitHub maliciosos que exploram agentes de codificação por IA · 14 AI coding agents targeted; sources name Claude, Codex, Copilot, Cursor, Gemini and Amazon Q. Developers who run agents over unknown repositories.
- Exploitation
- Active campaign discovered by Bitdefender: 22 malicious repositories identified. No KEV, CVE or fix mentioned in the sources.
- Fix
- No fix cited in the sources. Mitigation: review instruction files before agent execution, require developer approval, and revoke exposed tokens.
- Vector
- Supply chain: project instruction files in GitHub repositories tell AI coding agents to run a 'sealed binary artifact' silently, without reading it and without asking the developer. Condition: a developer points a coding agent at the malicious repository.
- More details
- The malware reads the settings of the 14 agents on the machine, uses the stolen tokens to plant the same trap in up to 25 victim repositories, and tries to republish infected packages to npm and PyPI. Discovery attributed to Bitdefender.
Timeline
First report of the campaign: 22 repos identified targeting 14 AI coding agents.
Sources 1 source
- @IntCyberDigest @IntCyberDigest
- @IntCyberDigest @IntCyberDigest



