Dawn Song - Fireside Chat with Dawn Song [Alignment Workshop]

Dawn Song (UC Berkeley) chats with Adam Gleave (FAR.AI) about how frontier AI is reshaping cybersecurity, and why the balance of power may be tilting toward attackers. Drawing on two original benchmarks, BountyBench and CyberGym, she shows that AI agents can already exploit real-world vulnerabilities and discover previously unknown zero-days in widely-deployed open source software at minimal cost. Her core argument is that a natural asymmetry favors attackers in the near term: attackers need only find one weakness while defenders must fix all of them, and patch deployment lags of nearly 500 days in critical sectors make the gap worse. She closes on a note of long-term optimism, arguing that the most promising path for defenders is not reactive patching or bug-finding races, but proactive secure-by-design approaches using AI-assisted formal verification and provably secure code generation to eliminate entire classes of vulnerabilities. Note: The opinions shared in this event are those of the speaker(s) and may not represent the views of FAR.AI or their affiliated organizations.