AI models detect software flaws faster than companies can fix them

Covered by 2 sources
Read onForbes
Advanced AI systems like Anthropic's Claude Mythos and Zhipu AI's GLM-5.3 are identifying software vulnerabilities at unprecedented speeds, outpacing the ability of companies to patch them safely. This rapid discovery, illustrated by cases such as PHP2Shell, is straining cybersecurity defenses and traditional vulnerability management processes.

What happened

AI models have become highly effective at detecting software vulnerabilities, identifying flaws at a much faster rate than before. In recent cases, advanced AI systems such as Zhipu AI's GLM-5.3 and Anthropic's Claude Mythos have demonstrated the ability to find and even exploit vulnerabilities in widely used software. One notable incident involved the rapid discovery of critical flaws in the PHP engine, which underpins many web applications. As a result, companies are struggling to keep up with the pace of fixing these newly found issues.

Key details

  • Zhipu AI's GLM-5.3 developed working exploits in 50 out of 410 attempts on a benchmark test.
  • Anthropic's Claude Mythos Preview achieved 56 working exploits on the same benchmark.
  • The PHP2Shell vulnerabilities (CVE-2026-109618 and CVE-2026-109619) affected the PHP engine and had a CVSS score of 9.2.
  • A study of 6,080 AI-generated patches found that 53.9% failed to resolve the original vulnerability or introduced new flaws.
  • A separate attack compromised LiteLLM version 1.107.0 by embedding malicious code in a Python package.

Why it matters

Security teams at organizations using popular software platforms now face a backlog of vulnerabilities that are being discovered faster than they can be safely fixed. This creates a risk that critical flaws may remain unpatched or that rushed fixes could introduce new problems. Companies must now allocate more resources to validate, prioritize, and deploy patches, increasing operational costs and complexity in maintaining secure systems.

Also covered by 1 source

More on this story

  1. ‘A failure of implementation’: AI played a key role in some outages—but not in the way you might think
    · 3 sources

More in Cybersecurity