@openai/codex-securityis a CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities in your code. Scan repositories, review changes, track findings over time, and run security checks in CI.
2 Likes
Prompt:
Why don’t we hear more about AI patching? Everybody would buy that so vulnerabilitiies would get fixed really fast
Reply from Gemini:
While every engineering team would buy a magic button that instantly fixes vulnerable code, the reality is that generative AI is currently much better at breaking things than fixing them.
Here is why automated remediation is hitting a massive wall:
1. It Introduces New Vulnerabilities
Recent 2026 research testing frontier models on complex vulnerabilities found that more than half of AI-generated patches are broken. The overall success rate for fully patching a flaw without creating new problems was less than a coin flip (47%). Even worse, in many tests, the AI didn’t just fail to fix the original bug—it actually introduced brand-new, exploitable flaws (like OWASP Top 10 vulnerabilities) into the codebase.
2. The “Band-Aid” Problem
When AI does manage to write a patch, it often creates what security researchers call “fragile guard code”. Instead of understanding and fixing the root architectural cause of a vulnerability, the model simply slaps a filter on the specific line of code that triggered the alert. This might pass an automated test, but it leaves other exploit paths wide open.
3. Context Decay
To patch a complex bug—like cross-site scripting or a kernel flaw—a developer needs to understand how data flows through the entire application. AI models have finite context windows. As they analyze massive codebases, they can lose context, forget how different modules interact, and start hallucinating fixes that technically compile but break the application’s actual business logic.
The Bleeding Edge: DARPA AIxCC
This is recognized as such a massive hurdle that the U.S. government is running a two-year, $29.5 million competition called the DARPA AI Cyber Challenge (AIxCC). The entire goal of the program is to figure out if autonomous systems can actually be trusted to discover and remediate vulnerabilities in real-world open-source software. Even among the best engineering teams in the world, getting an AI to reliably write production-grade patches without human oversight remains an unsolved problem.
Right now, if you deployed a fully autonomous AI patching agent into a production environment, you wouldn’t necessarily get a secure application—you would likely get a catastrophic system outage.
1 Like