Teaching AI to Find Real Vulnerabilities — Prof. David Brumley, Bugcrowd

Teaching AI models to hack requires a structured approach mirroring human learning, where tasks increase in difficulty from toy problems to hardened targets. Effective reinforcement learning environments must avoid simplistic crash-based grading, which encourages reward hacking and stunts model growth. Implementing an "audit task" framework—which evaluates precision and recall across multiple potential vulnerabilities—prevents models from fixating on single, easy bugs. Testing against high-value targets like Chrome’s V8 engine reveals that while many models can trigger crashes, only advanced systems achieve full sandbox escapes and arbitrary code execution. These findings demonstrate that successful cybersecurity training for AI relies on deterministic oracles that distinguish between various exploit primitives, ultimately moving beyond mere bug discovery to genuine, autonomous weaponization capabilities that rival the performance of elite human researchers.