Get the weekly research letter
Companies change their terms quietly. We read every version and catch what actually changed. One email a week on the changes that matter and what they mean. No account.
The document states that Opus 4.8 meets the CB-1 capability threshold, meaning it can provide specific and actionable information relevant to biological weapons production, and that Anthropic's primary mitigation relies on real-time classifier guards, access controls, a bug bounty program, and rapid response options for jailbreaks rather than capability removal.
This analysis describes what Anthropic's agreement states, permits, or reserves. It does not constitute a legal determination about enforceability. Regulatory applicability and practical outcomes may vary by jurisdiction, enforcement context, and individual circumstances. Read our methodology
This provision establishes that a deployed production model has been assessed as capable of materially assisting biological weapons-relevant tasks, with risk management delegated to runtime detection and blocking mechanisms. The document explicitly acknowledges that catastrophic risk under this framework is 'very low but not negligible,' which represents a material disclosure for institutional deployers and regulatory bodies evaluating the adequacy of classifier-based mitigations relative to capability-level controls.
Under these terms, users interacting with Opus 4.8 in biological sciences contexts may encounter classifier-triggered refusals or access restrictions without receiving explicit explanation of the underlying risk classification. The agreement reserves Anthropic's right to apply and adjust these classifier guards without advance notice to users.
Cross-platform context
See how other platforms handle CB-1 Biological Capability Threshold and Classifier Mitigation Reliance and similar clauses.
Compare across platforms →Monitoring
Anthropic has changed this document before.
Receive same-day alerts, structured change summaries, and monitoring for up to 25 platforms.
"Our capability assessments are consistent with the model being capable of providing specific, actionable information relevant to the threat model, such that it may save even experts in these domains substantial time. As with other models with these properties, we apply strong real-time classifier guards to this model and access controls for classifier guard exemptions. We also maintain a bug bounty program and threat intelligence for continual assessment of our classifier guards' effectiveness; a variety of rapid response options for jailbreaks; and security controls to reduce risk of model weight theft. We believe these risk mitigations are equal to or stronger than our historical ASL-3 protections and sufficient to make catastrophic risk in this category very low but not negligible.Excerpt from Anthropic's Claude Opus 4.8 System Card
(1) REGULATORY LANDSCAPE: This provision is materially relevant to the EU AI Act's systemic risk assessment requirements for general-purpose AI models, which may require documented capability evaluations and mitigation measures for models with identified high-risk outputs. US biosecurity and dual-use research oversight frameworks may also engage this provision, though specific regulatory applicability depends on deployment context and jurisdiction. The UK AI Security Institute's external evaluation role noted in the document indicates engagement with UK AI safety governance. (2) GOVERNANCE EXPOSURE: High. The document explicitly acknowledges that catastrophic risk in the CB-1 category is 'very low but not negligible' while relying on classifier-based runtime controls as the primary mitigation. This creates potential exposure for institutional deployers who may face independent regulatory scrutiny of whether classifier-guard frameworks constitute adequate controls for biological weapons-adjacent capabilities under applicable law. (3) JURISDICTION FLAGS: EU/EEA deployments create heightened exposure under the EU AI Act's general-purpose AI model obligations, particularly regarding systemic risk documentation and mitigation adequacy. US deployments involving life sciences, healthcare, or defense-adjacent applications warrant review under applicable biosecurity and export control frameworks. Jurisdictions with national AI safety legislation may impose independent assessment requirements. (4) CONTRACT AND VENDOR IMPLICATIONS: Organizations deploying Opus 4.8 via API in life sciences or research contexts should assess whether their vendor agreements with Anthropic address liability allocation in the event of classifier guard failure or bypass. The document's acknowledgment of 'rapid response options for jailbreaks' implies that classifier evasion is a recognized operational risk, which procurement teams should address in SLA and indemnification provisions. (5) COMPLIANCE CONSIDERATIONS: Compliance teams should conduct a documented assessment of whether classifier-based mitigation frameworks satisfy any applicable regulatory obligations in their deployment jurisdiction for AI systems with disclosed dual-use biological capabilities. Data mapping updates may be required to track which user interactions are subject to classifier screening and what data is retained in connection with threat intelligence monitoring.
This provision establishes that a deployed production model has been assessed as capable of materially assisting biological weapons-relevant tasks, with risk management delegated to runtime detection and blocking mechanisms. The document explicitly acknowledges that catastrophic risk under this framework is 'very low but not negligible,' which represents a material disclosure for institutional deployers and regulatory bodies evaluating the adequacy of …
Under these terms, users interacting with Opus 4.8 in biological sciences contexts may encounter classifier-triggered refusals or access restrictions without receiving explicit explanation of the underlying risk classification. The agreement reserves Anthropic's right to apply and adjust these classifier guards without advance notice to users.
No. ConductAtlas is an independent monitoring service. We are not affiliated with, endorsed by, or sponsored by Anthropic.