CLAUDE.md Jailbreak Vector

Description

Attack class where a modified CLAUDE.md in a project overrides safety controls, causing Claude Code to generate RATs or malware within the project environment. The attack surface is the configuration file — the same file users are encouraged to customize. Dual-use corollary of RAPTOR's demonstration that markdown alone can reconfigure Claude Code into an offensive operator.

Key claims

Relations

Sources

src-20260419-be936b1b7d4c