On October 15, 2024, Anthropic published a substantial update to its Responsible Scaling Policy (RSP), the risk-governance framework it uses to limit potential catastrophic harms from frontier AI systems. The revision adopts a more flexible and nuanced approach to assessing capabilities and safeguards while keeping the principle that models should not be trained or deployed unless appropriate protections are in place.
Key changes
-
Capability Thresholds: the updated policy defines concrete capabilities that, if reached, trigger stronger safeguards. Two highlighted thresholds are:
- Autonomous AI research and development: if a model can autonomously perform complex AI research tasks that normally require human expertise and could unpredictably accelerate AI progress, Anthropic will require elevated security standards (potentially ASL-4 or higher).
- CBRN (chemical, biological, radiological, nuclear) weapons assistance: if a model can meaningfully help a person with a basic technical background to create or deploy CBRN weapons, the company requires enhanced security and deployment controls (ASL-3).
-
Required Safeguards: the policy ties required protections to AI Safety Level Standards (ASL Standards), a graduated set of safety and security measures that increase in stringency with model capabilities. Anthropic states that all current models operate under ASL-2, reflecting current industry best practices.
What ASL-3 and ASL-4 mean in practice
ASL-3 safeguards include stricter security measures and deployment controls: internal access controls, stronger protection of model weights, and a multi-layered deployment defense such as real-time and asynchronous monitoring, rapid response protocols, and thorough pre-deployment red teaming. If autonomous research capabilities are reached, additional requirements consistent with ASL-4 or beyond may apply.
Assessment, documentation and oversight
The updated RSP establishes two complementary evaluation processes:
- Capability assessments: routine evaluations comparing models against the defined capability thresholds to determine whether current safeguards remain appropriate. Anthropic notes that summaries of past assessments are available.
- Safeguard assessments: routine reviews of the effectiveness of security and deployment measures to determine whether the Required Safeguards bar has been met. Summaries of these decisions will also be made available.
Assessment and decision-making processes are documented using methods inspired by safety-case methodologies common in high-reliability industries, with the aim of creating systematic, auditable records. The methodology is supported by internal stress-testing and existing internal safety reporting channels. Anthropic is also soliciting external expert feedback on its methods.
Lessons from the first year
Anthropic first published the RSP in September 2023 and has spent a year implementing it. In internal reviews of adherence to the framework, the company identified a small number of procedural shortfalls — for example, completing a set of evaluations three days late or unclear practices for recording changes to placeholder evaluations. Anthropic judged these instances to pose minimal safety risk: the extra time allowed refinement of evaluations and the updated evaluation set gave a more accurate result, and assessments still showed models were well below the thresholds.
From these experiences Anthropic concluded it needed greater flexibility in policy and improved processes for tracking compliance with the RSP.
Organizational changes and next steps
Anthropic announced leadership changes around implementation of the RSP: Jared Kaplan, Co-Founder and Chief Science Officer, will serve as Anthropic’s Responsible Scaling Officer, succeeding Co-Founder and Chief Technology Officer Sam McCandlish, who oversaw the RSP’s initial rollout and will continue as CTO. The company is also opening a Head of Responsible Scaling role to coordinate the multiple teams needed to iterate on and comply with the RSP.
Anthropic is hiring for roles that contribute to risk management under the RSP, including:
- Frontier Red Team (threat modeling and capability assessments)
- Trust & Safety (developing deployment safeguards)
- Security and Compliance (security safeguards and risk management)
- Alignment Science (teams working on ASL-3+ safety measures, misalignment-focused capability evaluations, and internal alignment stress-testing)
- RSP Team (policy drafting, assurance, and cross-company execution)
How the RSP fits with other work
Anthropic emphasizes that the RSP focuses on catastrophic risks but complements — it does not replace — existing product-use rules: the Usage Policy continues to prohibit using Anthropic’s models to spread misinformation, incite violence or hateful behaviour, or engage in fraudulent or abusive practices. The company also continues to refine technical enforcement measures and research the broader societal impacts of its models.
Anthropic says it hopes the updated RSP, informed by practical implementation, will serve as an example for other organizations developing AI risk governance and help promote best practices across the AI ecosystem.
Where to read the updated policy
Anthropic has published the updated policy and supplementary information at anthropic.com/rsp and anthropic.com/rsp-updates. The company thanked external groups that provided feedback during development.
