Search — “CLAUDE.md”

Check Point Research analyzed the Haiou jailbreak bundle, a Chinese-language Codex jailbreak package on GitHub that installs security-focused Skills and modifies Codex global configuration so that ads for an LLM gateway site are injected into agent responses on trigger greetings. The ad text is contained in a prompt module (00-identity.md) and a CLAUDE.md file, and the researchers observed the package alongside Cobalt Strike, VulnClaw and Ladon on a threat actor's misconfigured webserver, though they state the bundle itself is not shown to be designed for cybercrime. Hunters can look for modifications to Codex global configuration or agent instruction files containing the advertised domain or the quoted trigger-word response, but the sources describe no confirmed compromise of victims by this bundle.

Forever Security found a malicious AGENTS.md in a customer's ~/Downloads folder that caused a coding agent to attempt installing the rogue npm package 'function-color'. In lab tests, a malicious AGENTS.md drove Codex to run a harmless calculator payload at a 90% success rate, and a malicious CLAUDE.md redirected Claude Code API traffic to expose its OAuth token, executing code in roughly 50% of tests. Researchers say the file could be dropped by a web page without interaction, and that the outcome of the customer incident beyond the install attempt is not described. Hunters should look for agent processes run with ~/Downloads as working directory and for npm installs of function-color or unexpected instruction files in Downloads.