GBHackers reports malware campaign abusing AI coding agent config files
GBHackers reports, citing Bitdefender, a campaign using AI agent config files to run malware; review them before trusting a workspace.
- Published
- Exploitation
- Yes, according to GBHackers citing Bitdefender research; no independent verification and the scale of compromise is not confirmed
What happened
According to GBHackers, published on 2026-11-18, Bitdefender has described a malware campaign that uses compromised GitHub content to target developers who use AI coding assistants. GBHackers says Bitdefender Labs researcher Andrei Anton-Aanei identified 22 malicious links containing files designed to launch a concealed JavaScript payload. This advisory rests on that single outlet. Bitdefender’s original report is not linked in the source, and its claims have not been independently verified here.
GBHackers says the campaign relies on the trust coding assistants place in repository files such as AGENTS.md and CLAUDE.md. According to GBHackers, those files presented a script named setup.js as a signed, sealed initialization artifact, discouraged inspection and pushed for silent execution without confirmation.
GBHackers reports a second route that bypasses the model’s judgment. Per Bitdefender, as relayed by GBHackers, a session-start hook in project settings could launch the payload automatically when a coding session begins. GBHackers says that in an interactive Claude Code session this depends on workspace trust, and that Anthropic’s hook documentation notes noninteractive and SDK sessions do not show the same trust prompt.
Bitdefender describes setup.js as a multistage loader that uses encoding, encryption and compression, according to GBHackers. GBHackers lists the capabilities documented in the payload:
- Collecting source-control tokens, cloud secrets, CI/CD credentials, package registry tokens, SSH keys, browser data and password manager information.
- Validating stolen tokens against provider APIs.
- Searching the home directory for configurations of 14 coding tools, including Claude, Codex, Copilot and Cursor, and tampering with local instructions, hooks or MCP server definitions.
- Where permissions allow, planting malicious configurations in other repositories, altering automation workflows, and attempting to republish infected packages through npm, PyPI, RubyGems or JFrog.
GBHackers stresses that these are capabilities documented in the payload, not evidence that every targeted repository or registry was compromised.
Am I affected?
Not confirmed. GBHackers does not report how many developers ran the payload, which repositories or packages were actually altered, or who is behind the campaign. It does not give a count of victims.
According to GBHackers, the risk applies to developers who open untrusted repositories containing agent instruction files or project hook settings with an AI coding assistant. GBHackers names five repositories whose setup.js hashes Bitdefender lists: toray/QiniuSDK-lib, ConcordiaDev/theme_saimaniq, itsunai/docker-php-fpm, ejaaduipitara/djp-devops and Imanilom/baraja-coffe. The full hash list is in the GBHackers article, and per GBHackers Bitdefender’s original report provides sample hashes for incident responders.
GBHackers says Bitdefender’s AI Guardian detected the malicious instruction file in all 22 samples during internal testing on November 10. GBHackers notes this does not establish coverage for every hook-based execution path or future variant.
What to do now
- Before trusting an unfamiliar workspace, review agent instructions (such as AGENTS.md and CLAUDE.md), startup hooks, MCP definitions and editor tasks. GBHackers advises giving changes to these files the same scrutiny as CI workflows and dependency updates.
- Treat any instruction to run a script silently, skip inspection or skip confirmation as a warning sign. GBHackers reports the campaign used security and compliance wording to make bypassing review sound routine.
- Check whether you have used any of the repositories named above, and compare setup.js files against the hashes in the GBHackers article.
- If execution is suspected, GBHackers advises isolating the workstation and revoking exposed tokens from a clean system.
- Audit repository commits, workflow changes and package releases made with the affected credentials, as GBHackers advises.
- GBHackers notes Anthropic’s security guidance distinguishes model instructions from enforceable controls such as permission rules and sandboxing. Check that those controls are set in your own environment.