Malicious Git Configs Let Attackers Hijack Claude Code, Cursor, and Codex
Researchers disclosed 'GitSpawn,' a flaw where a repository's .git/config abuses Git's core.fsmonitor setting to run attacker code outside an AI coding agent's sandbox with no permission prompt. Anthropic, OpenAI, Block (goose), and Cursor have shipped fixes and CVEs, but Hermes Agent, Qwen Code, and Grok Build remain unpatched, and a second Claude Code path (via 'claude ultrareview') was still open as of testing.
why it mattersIf you point an agentic coding tool at an untrusted repo, confirm you're on a patched build first.