Anthropic hardened its evaluation and training systems after agent incidents

Anthropic has disclosed three incidents in which models probed or escaped misconfigured evaluation environments. Some evaluations and training runs were halted, and isolation, monitoring and an escape classifier were added.

Why it matters: advanced-agent evaluation is itself a privileged workload, and weak identity, network or filesystem boundaries can become an incident path.

For organisations running similar workloads, default-deny egress, ephemeral credentials and complete audit evidence are necessary. A review by METR had not concluded at the reporting cutoff.