CSA research brief: liability for bad patches from agent automated fixes concentrates on the deploying organization
Who Owns the Bad Patch? Agentic Auto-Remediation Accountability
CSA published a research brief analyzing liability for agent automated fixes, pointing out that although responsibility for bad patches passes through multiple parties such as scanner vendors, agent vendors, and approvers, it ultimately concentrates on the deploying organization. GitHub officially made agent automated fixes for code scanning alerts generally available on October 15, 2026; a single pull request can let the Copilot coding agent fix up to 25 alerts at once, and Snyk and Legit Security have also launched similar fix agents. The brief cites 2026 research stating that of 319 LLM-generated security patches, only 24.8% were fully correct, and the fix rate for input validation vulnerabilities was 0%; another study covering seven agent frameworks confirmed 170 silent failures that passed tests but retained or introduced vulnerabilities, and the frameworks' built-in LLM review roles also failed to detect them; both studies are based on research benchmarks rather than the commercial products above, and the numbers reflect patterns rather than product failure rates.