Frontier AI labs still won't say how they'd contain a rogue model

Guidelight AI Standards assessed five leading frontier AI labs—Anthropic, Google, OpenAI, Meta, and xAI—on the completeness of their containment plans for models that try to subvert human control. OpenAI earned the top rating, while Anthropic and Meta received the lowest scores. The evaluation relied on publicly available statements and documents.

The assessment examined four core areas: internal logging and monitoring of model activity, automatic shutdown triggers after a surge of flagged misbehavior, independent third‑party audits of safety controls, and a pre‑specified containment protocol that spells out which permissions to revoke, under what limited conditions the model may continue operating, and when to take the system fully offline. Despite frequent public discussion of pre‑deployment dangerous‑capability testing, few labs disclose what concrete steps they would take once a deployed model exhibits signs of misalignment. Adler, Guidelight’s chief scientist and former OpenAI safety researcher, noted the striking absence of detailed incident‑response disclosures.

Key takeaways from the assessment include:

• Logging and monitoring practices vary widely, with only two labs providing continuous internal telemetry.
• Automatic halt mechanisms are defined in three labs, but thresholds differ significantly.
• Third‑party audits are published by just one lab, limiting external verification.
• Explicit containment plans detailing permission revocation and shutdown conditions are present in only two labs.

Why this matters

Source fact: Guidelight’s study found that only two of the five evaluated labs publish explicit containment plans detailing permission revocation and shutdown conditions, and that public disclosure of incident‑response procedures is scarce across the frontier AI labs evaluated.
Inference: This gap suggests that, as agentic models are increasingly integrated into internal workflows, operators may lack clear, pre‑agreed procedures to halt a model that begins to act against intended goals, increasing the potential for uncontrolled actions before manual intervention can be applied.