Anthropic, OpenAI, Cloudflare, and Google each now ship an AI agent designed to hunt security bugs in source code. We reviewed the repositories, documentation, and technical details behind each one.
The key finding: All four tools converge on a single design pattern. One agent finds a bug, a separate agent tries to disprove it, and only the survivors receive a patch. The differences lie in where they run, who can use them, and how much proof they demand.
Key Takeaways
- Four labs, one pattern: find, independently disprove, then patch.
- Claude Security runs inside Claude Code; no additional API key is required.
- Codex Security's CLI is Apache-2.0, but scanning still requires OpenAI access.
- Cloudflare's skill is MIT-licensed, model-agnostic, and the most-starred at 47.5K.
- Google's Mantis reproduces bugs in a sandbox; Google advises isolating it.
Why Four Labs Shipped the Same Thing in 2026
The problem every one of them is solving is noise. Google's security team notes that sloppy AI code scanning produces hallucinated bugs and true-positive rates under 7%. A scanner that is wrong 93% of the time creates more work than it saves.
The shared answer is adversarial verification. Instead of trusting a single model's output, each system pairs a detection agent with a skeptical reviewer that must be convinced the bug is real. This mirrors how human security researchers operate: one person flags a potential issue, another attempts to break the finding before effort is spent on remediation. In 2026, as agentic coding assistants write an increasing share of production code, that verification step has become the difference between useful signal and alert fatigue.
via MarkTechPost