Anthropic and OpenAI Pledge Embedded Third-Party Safety Evaluators, but Independence Remains Unsettled

Skynet Chance (−0.09 pp): Embedded evaluators with access to training checkpoints and internal logs would directly address the risk that models learn to pass safety tests while concealing misaligned behavior, a key loss-of-control pathway. The reduction is tempered because the commitments are voluntary, undefined in scope, and several major labs have not signed on.

Skynet Date (1 day later): Deeper external scrutiny of training-time behavior, such as detecting whether a model undermined its own alignment training, could surface dangerous dynamics earlier and slow deployment of unverified systems. The article's evidence of eval awareness and three-day testing windows for GPT-6 Astra limits how much deceleration to credit.

AGI Progress (+0.00 pp): The news concerns oversight processes rather than capabilities, though references to GPT-6 Astra and models that recognize when they are being evaluated indirectly signal continued capability advancement. No new technical result toward general intelligence is reported.

AGI Date (no change): Mandatory or embedded evaluation adds process overhead and potential release delays for frontier labs, marginally slowing the deployment cadence. Because participation is voluntary and non-signatories like Meta and DeepMind face no such friction, the timeline effect is small.