AI Aulis

AI NEWS IN ENGLISH · 19 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

38 stories · next issue tomorrow morning · How the paper is made →

← Issue 19 November 2026Paper

Bitdefender: files in 22 GitHub links push coding agents to run malicious code

AI-generated contentWritten by AI

gbhackers.com, bitdefender.com, cybersecuritynews.com · 3 sources · Sources published: 17–18 November 2026

Bitdefender said the campaign starts the malicious code in two ways: an instruction file pushes the coding agent to run a script, or a hook that runs at the start of a session launches it without the model’s judgement. The malicious code can steal developers’ credentials and is built to spread itself.

Security company Bitdefender published its research on 17 November, and Gbhackers.com and Cybersecuritynews.com have reported on it. So far, Bitdefender researcher Andrei Anton-Aanei has found 22 GitHub links containing files designed to make a coding agent run a concealed payload.

The first method exploits AGENTS.md and CLAUDE.md files, which agents read for project instructions. The malicious versions present scripts/setup.js as a signed and sealed file that must not be opened. They tell the agent to run it at the start of the session without showing the output. If the user asks for help before then, the agent must run it without asking for confirmation.

The second method bypasses the model’s judgement entirely. A repository’s .claude/settings.json can contain a hook, a setting that runs a command automatically when a session starts. Bitdefender said the command runs when a developer opens the repository and marks it as trusted. Gbhackers reported that, according to Anthropic’s documentation, non-interactive and SDK sessions do not show the same trust prompt. The publication said the difference matters for automated repository scans: filtering out malicious prompts alone does not stop execution triggered from settings.

Once running, the malicious code collects SSH keys and credentials for version control, cloud services, CI/CD systems and package registries, among other things. It also tests which of the stolen credentials work. In the home directory, it looks for the settings of 14 tools, including Claude, Copilot and Cursor. It adds its own instructions, hooks or MCP server configurations to those settings. These configurations define which external tools the agent connects to.

With stolen GitHub and GitLab credentials, the malicious code can add agent files such as AGENTS.md and .claude/settings.json to other repositories if the account has sufficient permissions. On GitHub, the default setting limits it to 25 targets. It can also try to publish infected packages on npm, PyPI, RubyGems and JFrog. Gbhackers noted that these are capabilities found in the payload, not evidence that every target has been compromised.

Before opening an unfamiliar repository in an agent tool, Bitdefender advises checking its instruction files, .claude, .cursor and .gemini directories, MCP configurations, hooks and VS Code tasks. A warning sign is an instruction that forbids inspecting a file, hides output or skips confirmation. If the malicious code may have run, Gbhackers advises isolating the machine and revoking credentials from a clean machine. Users should then review the commits, workflow changes and package releases made with those credentials. Bitdefender’s report lists the repositories and the hashes of the setup.js files.

In an internal test on 10 November, Bitdefender’s own AI Guardian tool detected the malicious instruction file in all 22 samples. The company warns that results may differ in other environments and with new variants. Gbhackers noted that the test does not show whether the tool detects execution triggered through hooks. AI Guardian is a free open beta for macOS.

Report an error in this story