Anthropic has introduced its Responsible Scaling Policy (RSP) to manage the risks of advanced AI systems. The RSP features AI Safety Levels (ASL), inspired by biosafety standards, categorizing AI based on potential risk from ASL-1 to ASL-4. Current models, such as Claude, are classified as ASL-2, indicating early hazardous capabilities. The policy aims to minimize catastrophic risks while promoting beneficial AI applications, halting model training if safety standards can’t be met. Continuous evaluation, approved by the board, ensures rigorous safety measures akin to those in automotive and aviation industries. The RSP is a dynamic framework, adapting to new AI developments.

Anthropic has launched a Responsible Scaling Policy (RSP) to mitigate risks amid AI evolution. The policy introduces an AI Safety Levels (ASL) framework, inspired by biosafety standards, with five escalating levels of safety measures to manage AI’s catastrophic risks. Presently, most Anthropic models, including Claude, are classified under ASL-2. The policy ensures higher safety protocols and a commitment to halt advanced model training if safety cannot be upheld. The RSP, endorsed by Anthropic’s board, aims to balance risk mitigation with beneficial AI applications. Supported by ARC Evals, the RSP provides guidelines for safe AI deployment across various stakeholders.

Anthropic has launched an advanced “frontier threats red teaming” initiative to enhance AI safety against national security risks. This program focuses on adversarial testing, particularly in biology and cybersecurity, involving domain experts to identify and mitigate potential AI threats. CEO Dario Amodei emphasized these concerns during Senate testimony, aligning with a White House commitment for rigorous AI testing. Initial findings show current AI models can accelerate harmful activities, though mitigations like classifier-based filters are being integrated. Anthropic aims to scale this work, encouraging broader collaboration and disclosure processes to bolster overall AI safety and address long-term risks.

Anthropic has introduced the Responsible Scaling Policy (RSP) to manage risks from advanced AI systems. The RSP uses a tiered AI Safety Levels (ASL) approach, akin to biosafety levels, to establish safety standards based on potential catastrophic risks. ASL-1 covers negligible risk systems, while ASL-2 includes models with early dangerous capabilities. ASL-3 addresses higher risk systems, with ASL-4 and above yet to be detailed. The policy aims to foster safe AI innovation without halting advancements. Approved by Anthropic’s board, the RSP emphasizes rigorous pre-market testing and ongoing adjustment to rapid AI developments. Full policy details are available for stakeholders.

Anthropic has introduced its Responsible Scaling Policy (RSP) to address risks associated with advanced AI systems. The RSP uses an AI Safety Levels (ASL) framework, akin to U.S. biosafety standards, categorizing AI systems by their catastrophic risk potential. Current large language models, like Claude, are categorized as ASL-2, requiring higher safety but not disrupting existing products. Higher ASL levels demand stricter security and operational standards, including comprehensive adversarial testing. ASL-4 is under development for more autonomous and potentially misused models. The policy aims to balance mitigating catastrophic risks while encouraging beneficial AI applications. ARC Evals provided significant input for the RSP.