RSP 3.0
- The clause is changed. Anthropic will only halt development on AI systems if they are ahead, and there is enough material to halt development
AI Safety Levels
- ASL-1: small models with no meaningful catastrophic risk
- ASL-2: models that show early signs of dangerous abilities, but uses all information available to humans
- ASL-3: models that have substantial risk and show low-level autonomous capabilities
- ASL-4+: Systems that can cause catastrophic misuse and have high autonomy
RSP Safeguards
Anthropic will:
- Use Constitutional Classifiers to prevent jailbreaks
- Detect jailbreaks when they happen
- Continuously redteam their AI agents to prevent jailbreaks
RSP Evaluations
Rule-out Evaluations
Designed to check if model falls below a specific capability threshold
Rule-in Evaluations
Designed to check if model has gone above a capability threshold
Evals
- RSP is designed to assess model capabilities in catastropic risk against CBRN risks, autonomy and cyber domains
- ASL-3 evals check if the model has capability to obtain and deploy CBRN weapons
- ASL-4 evals check if model is ablle to create novel and new CBRN weaponry
- Both evaluated with skill-testing questions, and long-form task based agent evals
- Models provided Agent Harness with software tools, knowledge based tests gave models search and research tools
- Claude Opus 4.5 has extensive information on biology and biosecurity, but performs below on short-horizon biology computation tasks
- No evals on chemical risks
- Relies on NNSA (national security administration) for evaluations on nuclear and radiological risks
- LAB-Bench to measure ability of model to comprehend and reason about figures in biology papers
- SWE-bench to test if model can solve real-world github issues
- AI research evaluation suite 1 and 2 to test if models can do AI/ML research