AI fails to properly patch software vulnerabilities 74% of the time

Frontier AI models are failing to reliably fix software vulnerabilities, according to a new study finding that automated patches were unsuitable 74% of the time. The research reveals a dangerous gap between an AI's ability to identify security flaws and its capacity to resolve them safely.

Conducted by Off-By-1-Labs, a security research team within 1Password, the study—titled 'FLAWED' (Fix-Like Artifacts With Embedded Defects)—tested several large language models (LLMs), including Claude and a Codex-based agent. Researchers performed 6,080 patch attempts across six recently disclosed open-source vulnerabilities, utilizing nine different crafted prompts for each bug. The results revealed that only 26% of the generated patches were actually usable.

According to the data, 53.9% of the attempts resulted in the LLM failing to create a patch entirely, introducing a new bug, or both. An additional 21% of the patches successfully fixed the original vulnerability but inadvertently altered the application's intended behavior. These findings contradicted the researchers' initial hypothesis, which predicted a success rate of approximately 67% based on the volume of open-source code available in AI training sets.

The Risk of False Security

These findings are significant as organizations increasingly integrate AI into cybersecurity workflows. The study warns that over-reliance on automated patching can create a false sense of security, as AI-generated fixes often appear correct on the surface while containing embedded defects.

Off-By-1-Labs researchers noted that while LLMs excel at discovering a wide range of vulnerabilities, they are currently only effective at patching a narrow subset of them. This disparity means that while AI can accelerate the discovery of flaws at scale, it cannot yet be trusted to remediate them without rigorous human oversight, as it may introduce new vulnerabilities into production environments.

Shifting the AI Strategy

Given these limitations, the research suggests a pivot in how security teams utilize AI tooling. Rather than trusting AI to write the final code for a fix, the study argues that the technology is better suited for the early stages of the security lifecycle.

Keith Hoodlet, head of Off-By-1 Labs, stated that human defenders and AI tooling should instead focus on vulnerability discovery and triage. According to Hoodlet, this approach helps defenders determine where the most impactful bugs exist within their codebase, leaving the critical task of patching to human experts who can ensure the stability and security of the application.