Definition
The leading edge of general-purpose AI models — systems whose capabilities meaningfully exceed the prior state of the art on reasoning, coding, cyber-security, or other safety-relevant tasks. A term used by frontier labs (Anthropic, OpenAI, Google DeepMind) and by policy bodies (UK NCSC, US AI Safety Institute, EU AI Act) to scope additional governance attention to the most capable models.
Why It Matters for the Newsletter
Frontier AI is the category within which the wiki’s AI-state-conflict (Anthropic vs. US Department of Defense), AI-liability (AI Liability), and AI-security threads converge. The category matters because both governance and access controls are organized around it — models below the frontier are treated as commodities; frontier models are increasingly subject to gated release, DoD interest, and now (per Claude Mythos Unauthorised Access — BBC) vendor-permission breach risk.
Evidence & Examples
- Anthropic maintains gated release for its Mythos cyber-security model because of its reported vulnerability-exploitation capabilities; OpenAI has a parallel capable model, GPT 5.4 Cyber. Claude Mythos Unauthorised Access — BBC
- UK NCSC head Richard Horne (CyberUK 2026): “frontier AI is rapidly enabling discovery and exploitation of existing vulnerabilities at scale”
- Frontier AI is US/China-centric; UK and most allies rely on commercial access, with no control over training or release (Horne, in Claude Mythos Unauthorised Access — BBC) — a sovereignty-by-proxy arrangement
Tensions & Counterarguments
- “Frontier” is a moving target and self-designated by labs — useful as a governance abstraction, weaker as a technical boundary.
- Gated release policy depends on vendor and contractor access hygiene; the Mythos incident shows the gate is only as strong as the weakest third-party environment.
- Non-frontier open models (including fine-tunes of earlier frontier models) may already reach frontier capability on narrow tasks, making the frontier/non-frontier distinction less meaningful for risk assessment.
Related Concepts
- AI Sovereignty — frontier AI geography shapes sovereignty
- AI Liability — criminal/civil liability is currently being drawn around frontier-model outputs (Florida attorney general announces criminal investigation into OpenAI)
- Embodied AI — a frontier in physical-world task performance
- Reinforcement Learning — central training paradigm
Key Sources
- Claude Mythos Unauthorised Access — BBC — Mythos gating, NCSC framing, US/China concentration
- Sony Ace Table Tennis Robot Beats Human Pros — AP — “ChatGPT moment for robotics” framing