1-minute read of a 5 min report

Gemini and ChatGPT Recommended Malicious GitHub Skills in Tests

By

1-minute read

What a defender needs to know

Island mapped about 7,600 malicious FakeGit repositories, including more than 800 posing as AI skills or MCP servers. Credible READMEs and lookalike profiles directed users to ZIP files that installed SmartLoader and then StealC. More than 600 campaign listings appeared across public AI capability catalogs.

In controlled tests, Claude Code, Gemini and ChatGPT independently surfaced campaign-linked repositories while searching for skills or MCP servers. Some runs recognized warning signs, while others repeated attacker-written setup instructions. The tests do not measure real-world compromise rates, but they confirm that agents can carry a lure from ordinary capability discovery to a user.

Use an approved catalog for company-connected skills and MCP servers. Record the reviewed repository, commit, version and package hash. Test unfamiliar software on an isolated system without live sessions or developer credentials. Require external policy approval before an agent executes downloaded programs or setup commands. If SmartLoader may have run, isolate the endpoint and revoke browser sessions, OAuth grants, API tokens and cloud or developer credentials.

Primary sources