Black Hat Asia 2024 Keynote: AI's Rapid Evolution in Cybersecurity and Offensive Capabilities

The Black Hat Asia 2024 keynote addressed the rapid evolution of AI in cybersecurity, particularly offensive capabilities. Jeff Moss and Ari Herbert-Voss highlighted how large language models (LLMs) like Anthropic’s Mythos and OpenAI’s GPT-5.5 are scaling superlinearly—doubling model size, compute, and data yields fourfold capability improvements—accelerating vulnerability discovery from months to hours. LLMs excel at finding shallow, high-volume bugs (e.g., 400+ tier 1/2 crashes in OSS-Fuzz evaluations) but struggle with state-dependent issues like concurrency or deep exploitation. While attackers leverage LLMs for volume, defenders must filter outputs, as reliability and targeting remain inconsistent. Open-source models and economic pressures are democratizing access, with commercial and criminal actors neck-and-neck in adoption. The talk noted AI’s limitations in debugging and temporal analysis, though detailed error logs improve performance. Black Hat also announced its expansion to India (October 27–30, 2024).