Jessie A Ellis
Oct 31, 2025 11:40
Anthropic has activated AI Safety Level 3 standards to bolster security and deployment measures, particularly against CBRN threats, with the launch of Claude Opus 4.
Anthropic, a leading AI research company, has announced the activation of its AI Safety Level 3 (ASL-3) Deployment and Security Standards. This move is part of the company’s Responsible Scaling Policy (RSP) and coincides with the launch of Claude Opus 4, according to Anthropic.
Enhanced Security Measures
The ASL-3 Security Standard introduces advanced internal security measures designed to prevent the theft of model weights, which are crucial to the AI’s intelligence and capability. These measures are particularly focused on countering threats from sophisticated non-state actors. The deployment standards aim to limit the risk of the AI being misused for the development or acquisition of chemical, biological, radiological, and nuclear (CBRN) weapons.
Proactive Implementation
Although it has not been conclusively determined that Claude Opus 4 requires ASL-3 protections, the decision to implement these measures was made proactively. This precautionary step allows Anthropic to test and refine its security protocols in response to the evolving capabilities of AI models. The company has ruled out the necessity of ASL-4 standards for Claude Opus 4 and ASL-3 for Claude Sonnet 4.
Deployment and Security Focus
The ASL-3 Deployment Measures are specifically tailored to prevent the model from aiding in CBRN-related tasks. These measures include limiting “universal jailbreaks,” which are systematic attacks that circumvent security guardrails to extract sensitive information. Anthropic’s approach includes making the system more resistant to jailbreaks, detecting them as they occur, and iteratively improving defenses.
Security controls focus on protecting model weights with over 100 different security measures, including two-party authorization for access and enhanced change management protocols. A unique aspect of these controls is the implementation of egress bandwidth controls, which restrict the flow of data out of secure environments to prevent unauthorized access to model weights.
Continuous Improvement
Anthropic emphasizes that the implementation of ASL-3 standards is a step towards ongoing improvement in AI safety and security. The company continues to evaluate the capabilities of Claude Opus 4 and may adjust its security measures based on new insights and threat landscapes. Collaboration with other AI industry stakeholders, government, and civil society is ongoing to enhance these protective measures.
Anthropic’s comprehensive report provides further details on the rationale and specifics of these newly implemented measures, aiming to serve as a resource for other organizations in the AI sector.
Image source: Shutterstock
