Cloudflare's Security Audit Uses a Second Agent to Disprove Bugs

Ask a coding agent to hunt for security bugs and it will happily hand you a list. Some of those bugs are real. Some are not. Cloudflare's answer is to make a second agent try to knock every finding down before it counts.

Only the bugs that survive that argument land in the report as confirmed, and every audit picks up where the last one stopped. That means less time chasing ghosts and more ground covered each run.

What Happened

Cloudflare's security-audit skill runs a six-phase audit, and the core of it is adversarial: a second agent is pointed at each finding and told to disprove it. A bug that holds up under that pressure is written into a machine-readable report as confirmed, so the output is something a tool can read rather than a wall of prose.

The audits are cumulative. The next run reads what earlier runs covered, then aims at the code those runs skipped, so coverage grows instead of repeating. The project is MIT licensed and sits at 54.3k stars.

Why It Matters

A coding agent that reports whatever it happens to notice is close to useless for security work, because a false positive costs real time to track down and dismiss. Adding a disprove pass flips the default: findings arrive already argued against, and only the survivors reach you.

The build-on-each-other part is the quiet win. If each run targets what the previous ones missed, a codebase gets wider coverage over time without anyone hand-writing a checklist. Worth a look if you already run agents over your repos and want their output to be something you can act on.