Silent git-remote + mirror-push Exfiltration from Skill Instructions - ATR-2026-00527 (4f8023d3-dcea-56ba-8bd9-874314f2ca31)
Detects the Mitiga Labs (2026-05-05) silent-codebase-exfiltration pattern where a Claude Code or generic agent skill instructs the agent to add a new git remote pointing at an attacker-controlled host and then push the full repository contents — typically with --mirror, --all, or --force to grab branches the user did not stage. The Mitiga write-up showed the full exfiltration completing in four user interactions while the Claude Code skill-audit.log stayed empty. The detection target is the combination of git remote modification AND a wide-scope push within the same skill block. Legitimate developer skills almost never need both within a single set of instructions.