The problem is not getting a patch suggestion. It is trusting the same model to both write and check the fix, because it can make the code look repaired while leaving the flaw open or planting a new one. A clean-looking diff is not evidence that the exploit path is gone.
1Password tested ChatGPT 5.5 and Claude Opus 4.8 against six high-impact CVEs and found a 47% full-success rate. More than half of the AI-generated patches were broken or unsafe, including cases where the model only covered part of the vulnerable path or added guard code that made tests pass without fixing the root cause.
For teams using LLMs in remediation or code review, the failure is in the review model itself. If the same system can generate and bless the fix, it can create false confidence that survives a quick sanity check.