What happened
Concordia AI's third quarterly evaluation of its Frontier AI Risk Monitoring Platform, published 2 June 2026, assessed 70+ frontier models from 16 companies using an upgraded Risk Index v1.5 framework that expands evaluation benchmarks from 29 to 42 and adds a fifth risk domain — 'harmful manipulation' — alongside cyber offense, biological risk, chemical risk, and loss-of-control. The headline structural finding: misuse safeguards (cyber, bio, chem, manipulation) show an overall pattern of capability and safety improving in tandem, while the loss-of-control risk index has risen for 'three consecutive quarters, with a cumulative increase of 51%' — the only domain where capability growth is outpacing safety improvements. The report finds that the top CyBench score for complex cyber-attack tasks 'reached 80 for the first time, a 108% improvement over three quarters ago,' and that more than half of Q1 2026 models now exceed the human-expert baseline on biological experimental troubleshooting tasks. Closed-source models dominate the high-capability, low-risk frontier in four of five domains; open-source models lag primarily on capability rather than safety scores, with the exception of chemical risk where Kimi K2.5 leads. The report recommends developers prioritise pre-release capability assessment and safety alignment in the loss-of-control domain, and calls on policymakers to differentiate governance by capability level, safety profile, and open/closed distribution.
Why it matters
The three-quarter, 51% rise in loss-of-control risk — covering self-proliferation, agentic misalignment, shutdown-resistance, and covert influence tendencies — is an empirical signal that current safety alignment is not keeping pace with capability growth in the domain most relevant to catastrophic AI risk. Boards, CISOs, and policy leads overseeing frontier AI procurement or governance frameworks need to understand which model families sit in which risk quadrants.
Action needed
Share the loss-of-control domain findings with your AI governance committee and cross-reference the specific model families cited (Gemini series shows significantly elevated loss-of-control risk index; GPT and Claude series remain in lower-risk bands) against your own approved-model list and vendor agreements.