AnthropicAnthropicAnnouncement
Activating AI Safety Level 3 protections
Summary
Anthropic has activated AI Safety Level 3 (ASL-3) protections for its Claude Opus 4 model, a higher standard than the ASL-2 baseline used for all previous models. The decision was proactive, made before determining whether the model’s capabilities actually require the enhanced measures. ASL-3 deployment measures focus narrowly on preventing the model from assisting with extended, end-to-end CBRN weapons workflows. The approach includes Constitutional Classifiers that monitor inputs and outputs, a bug bounty program, and threat intelligence partnerships. Security controls include over 100 measures such as two-party authorization for model weight access and egress bandwidth controls that limit outbound data flow to prevent weight exfiltration.
Context
Anthropic’s Responsible Scaling Policy defines Capability Thresholds that require higher AI Safety Level Standards when models approach dangerous CBRN capabilities. All previous models operated under ASL-2, which includes training models to refuse CBRN requests and defenses against opportunistic weight theft. ASL-3 requires stronger defenses against sophisticated non-state attackers.
Approach / What changed
Anthropic implemented the ASL-3 Standard for Claude Opus 4, which includes deployment measures (Constitutional Classifiers that monitor inputs/outputs for CBRN-related content, a bug bounty program, offline classification, and threat intelligence partnerships) and security controls (over 100 controls including two-party authorization for model weight access, binary allowlisting, and egress bandwidth controls that limit outbound data flow to make weight exfiltration difficult).
Takeaways
ASL-3 deployment measures are narrowly focused on preventing models from assisting with extended, end-to-end CBRN workflows, not on blocking commonly available single pieces of information like chemical formulas.
Constitutional Classifiers use synthetic data to train real-time guard classifiers that monitor inputs and outputs, substantially reducing jailbreaking success while adding only moderate compute overhead.
Egress bandwidth controls leverage the large size of model weights by limiting outbound network traffic, making exfiltration very difficult even if an attacker has otherwise compromised internal systems.
FAQ
What are Constitutional Classifiers?
Constitutional Classifiers are real-time classifier guards trained on synthetic data representing harmful and harmless CBRN-related prompts and completions. They monitor model inputs and outputs and intervene to block a narrow class of harmful CBRN information, adding only moderate compute overhead to normal operations.
How do egress bandwidth controls protect model weights?
Egress bandwidth controls restrict the rate of outbound network traffic from secure computing environments where model weights reside. Because model weights are substantial in size, limiting bandwidth makes it very difficult for an attacker to exfiltrate the weights before detection, even if they have otherwise compromised the systems.
Will ASL-3 protections remain permanent for Claude Opus 4?
Anthropic may remove or adjust the ASL-3 protections if they conclude that Claude Opus 4 has not surpassed the relevant Capability Threshold. The RSP allows deploying under a higher standard than needed, and they will continue evaluating the model's CBRN capabilities.
Source and freshness
- Source
- anthropic.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.