How an AI‑Powered Loop Hunt Discovered Over 200 Real Bugs in Six Months

1. From One‑Time Scan to Never‑Ending Loop

Most static analysis tools perform a single pass over the code and stop, giving a false sense of completion. Researcher dinosn encapsulated his vulnerability‑hunting methodology into a Claude Code skill called raptor-loop-hunt , adopting the “Karpathy automated research” principle of keeping the AI in a perpetual “loop‑hunt” state.

2. Three Elements of the Loop Hunt

The core of raptor-loop-hunt is a generate → judge → verify cycle, implemented with three concrete mechanisms:

Multi‑level attack‑surface scanning : The tool starts from a system‑wide view and drills down to individual functions, analogous to inspecting a building floor by floor.

Adversarial verification : After a candidate vulnerability is generated, the system questions its own output by returning to the source code and performing an adversarial check, dramatically reducing false positives.

Evidence‑driven conclusions : Every confirmation, rejection, or downgrade of a finding must be backed by concrete evidence, ensuring that severe issues cannot be silently downgraded.

3. 200+ Bugs in Six Months

From December 2025 to July 2026 the project reported:

200+ verified vulnerabilities.

40+ real code repositories spanning C network daemons, message brokers, routing stacks, web applications, language runtimes, and media libraries.

The author disclosed only titles and categories of the bugs, withholding PoCs to demonstrate effectiveness without providing usable exploits.

4. How to Use

Option 1: Claude Code plugin (recommended)

/plugin marketplace add dinosn/raptor-loop-hunt
/plugin install raptor-loop-hunt@raptor
/.reload-plugins

Then describe a task, e.g., audit ./src for security.

Option 2: Install as a personal skill

git clone https://github.com/dinosn/raptor-loop-hunt ~/.claude/skills/raptor-loop-hunt

Invoke with /raptor-loop-hunt.

Option 3: Direct prompt

find every bug in this service
find anything exploitable

The skill parses the request and launches the automated workflow.

5. Why the Tool Matters

raptor-loop-hunt transforms a researcher’s mental methodology into a reusable artifact, offering three practical benefits:

Independent researchers can boost audit efficiency.

Enterprise security teams can adopt it for continuous code review.

It adds a concrete engineering path for AI‑augmented security capabilities.

It does not aim to replace human security engineers; instead, it amplifies the human strengths of relentless iterative thinking and adversarial validation.

Project URL: https://github.com/dinosn/raptor-loop-hunt