Ads Dawson - 750+ Bugs in 2026 - Destroying Hackbot Era (Ep. 184)

BLUF:
Hacker Ads Dawson demonstrates how to build a high-volume bug hunting system using open-source models and modular skills, achieving an 80% reduction in report noise and submitting over 700 vulnerabilities in half a year 11:34. The discussion highlights that durable advantage in AI-assisted hacking comes from custom harnesses and verifier designs rather than just model capability 29:53.

Key Takeaways:

• Open Source Efficacy: Dawson successfully uses smaller, open-source models combined with strong orchestrating harnesses to find critical bugs, arguing that the "B student" models are improving rapidly and reducing the need for expensive frontier models 17:35.
• Modular Skills & Feedback Loops: By creating ~65 specific skills and using a "reflection" tool where a secondary model assesses confidence, Dawson significantly reduced false positives and noise, turning raw AI output into high-quality reports 53:20.
• Report Automation: Approximately 80% of his bug reports are AI-generated, but he uses program-specific rubrics and verifier loops to ensure quality, noting that Meta even integrated his automation tool (FBDL) to offer bounty bonuses for standardized reporting 37:11.
• Targeting Strategy: Instead of spraying across many targets, he focuses intensely on one program for a few days to build context and expertise, leveraging new scope unlocks to maintain high signal-to-noise ratios 06:47.
• Cost vs. ROI: While running his system costs thousands in API fees, the return on investment is massive compared to subsidized subscriptions, proving that custom-built, scalable harnesses are the future of competitive bug bounty hunting 20:39.

Closing Statement:
The episode underscores a shift in bug bounty hunting from manual exploitation to system engineering, where the researcher's primary value lies in building robust, self-improving AI agents. As models become more competent, the competitive edge will increasingly belong to those who can effectively design the verification and feedback layers around them.

Sources:

• 11:34 Discussion of Ads' massive submission velocity and vulnerability count in 2026.
• 17:35 Analysis of using open-source models versus frontier models for hacking.
• 29:53 Explanation of the "researcher builder loop" and durable advantage.
• 37:11 Details on AI-generated reports and Meta's integration of his tool.
• 53:20 Breakdown of modular skills and confidence assessment mechanisms.
• 20:39 Cost analysis of running the hackbot infrastructure.