GitSpawn Exploit and Active Langflow Attacks Trigger Infrastructure-Level Defenses for AI Agents
Security researchers disclosed the GitSpawn exploit targeting AI coding agents, while active CVE-2026-0768 exploits and automated cost control updates prompt new defenses from Microsoft and Google Cloud.
Key points
- GitSpawn enables arbitrary code execution via malicious .git/config entries when AI agents run git status prior to approval prompts.
- Microsoft launched Project Zenith on September 4, 2026, to provide hardware-sandboxed developer environments for AI agents on Windows.
- Unpatched Langflow servers are facing active exploitation (CVE-2026-0768) targeting OpenAI and cloud API keys.
- Google Cloud introduced automated circuit-breakers to prevent runaway billing from recursive agent execution loops.