Four AI Vulnerability Hunters on One Stage: Under 7% True Positives—Why Are Security Teams Still Willing to Use Them?

In 2026, four labs each released an AI agent specifically designed to find security vulnerabilities in source code: Anthropic, OpenAI, Cloudflare, and Google. They did not each present a different methodology; their workflows are almost isomorphic: find first, then disprove, and finally patch.

This detail is more important than "AI can find vulnerabilities." Traditional source code scanning relies on rules and static analysis. Security teams are not short on alerts; they are short on alerts that can be validated, reproduced, and closed. AI agents read code, propose candidates, then try to disprove themselves, and finally generate patches—hitting exactly this pain point.

Google offered a sobering number: with sloppy AI code scanning, the true positive rate is below 7%. Note the scope: Google was talking about sloppy AI code scanning, not all AI security tools. 7% shows that candidate generation is cheap and validation is expensive. What security teams really need to buy is lower false positives and lower manual review costs, not another ten thousand alerts.

Cloudflare released its skill under the MIT license on June 18, 2026, with 47.5K stars, making it the most closely watched of the four. It reported 20,799 raw candidates, of which about 12,057 were validated. Another number is more practical: after using better recon context, Cloudflare reduced its first-round rejection rate from 40% to 11%. With more complete context, the agent guesses less, and early candidate quality improves.

Google also mentioned that its summary tree cut token costs by more than 85%. For ordinary readers, this means AI reading large codebases is expensive, and tokens are a real cost. Only by compressing context well can these tools enter daily workflows.

In 2026, at least 12 more agent-based vulnerability tools were released or made public. The field is already crowded. The fact that all four share a find, disprove, patch loop shows that the competitive point is not "discovery," but disproving and patch quality. Generating ten thousand vulnerability reports is not impressive; being able to prove that most of them are false positives and then provide mergeable patches is the reason security teams are willing to keep these tools around.