Anthropic's Pilot Sabotage Risk Report (alignment.anthropic.com)

Anthropic released a pilot "Sabotage Risk Report"—a practice run in advance of future Responsible Scaling Policy obligations—assessing misalignment risks from deployed models as of Summer 2025, primarily focusing on Claude Opus 4. The core conclusion is that sabotage risk (misaligned autonomous actions that could materially contribute to later catastrophic outcomes) is very low but not zero. Anthropic says it is moderately confident Opus 4 lacks consistent, coherent dangerous goals and does not have the capabilities to reliably plan and execute complex, undetected sabotage. The report treats sabotage as a distinct high-capability risk (separate from hallucination), enumerates several low-probability threat models, documents existing formal and informal safeguards (monitoring, training interventions), and acknowledges gaps that require further mitigation. The exercise is significant because it pioneers an affirmative-case style assessment—unprecedented, Anthropic argues—for present-day model autonomy and establishes a template for how frontier labs might demonstrate acceptable misalignment risk before scaling. The report underwent internal Alignment Stress-Testing and independent METR review; both endorsed the overall risk judgment while flagging weaknesses in evidence and argumentation. Anthropic reports concrete follow-ups (revised threat models and training changes to improve faithfulness and monitorability) and frames this as an iterative pilot: useful transparency and a partial blueprint for the community, but not a final safety certification.