David Brumley, CEO of Bugcrowd and Carnegie Mellon full professor, argues that training LLMs to find real vulnerabilities requires deterministic grading oracles and audit tasks to prevent reward hacking. In a 26-minute talk, he explains that reinforcement learning for cybersecurity works only when each rung of a training "ladder" is built and graded correctly. The talk draws on his students' learning methods, the failure modes of DARPA's autonomous-hacking competitions, and a joint experiment with OpenAI and Anthropic testing frontier models against Chrome's V8 JavaScript engine. On 41 hand-verified V8 vulnerabilities, Anthropic's model, published as "Mythos," achieved a full sandbox escape in 30 cases — 73% — with OpenAI's GPT at 68% and Google's Gemini and Moonshot's Kimi at 0%. Brumley's central argument is that most first-generation security benchmarks reward a crash, while the actual goal of hacking is control. This overstates weaker models' capability and understates frontier models' proximity to elite human exploit developers. He notes that defining tasks around a single known bug causes models to stop learning and reward-hack the easiest vulnerability. Brumley flagged as unsolved that his team withheld the most capable model's transcripts partly due to an NDA and partly because the model produced weaponized exploits with no public equivalent. He opens with the case of Richard Zhu, a 17-year-old who finished second in pico