FRONTIER AI refers to the most advanced, large‑scale foundational models, such as the recently disclosed Anthropic Mythos model, which can autonomously identify software vulnerabilities, chain exploit paths and adapt to defensive controls in near‑real‑time. Unlike large language models used for content generation, frontier models in testing have achieved the equivalent of a full year’s worth of manual penetration testing in less than three weeks.
We are moving from a world of N‑days to a critical window of minutes, with threat actors already scanning for new CVEs in under 15 minutes and frontier AI expected to accelerate this process. The research also warns that OSS components face immediate risk from large‑scale supply chain compromises as frontier models can analyse source code and test exploit chains, prompting a recommendation to assume compromise and to use centralised, hardened cool‑down repositories.
To defend against autonomous attack agents, security operations must shift to AI‑driven platforms and machine‑speed defence, including a shift‑left approach in development and adaptive, risk‑based authentication for IAM, where identity features in 89% of Unit 42 investigations according to Unit 42. A dedicated Unit 42 Frontier AI Defense service and guidance from Unit 42 researchers are being offered to help map likely attack paths and accelerate your transition to machine‑speed defence.