Residual Safety Risks and Jailbreak Acknowledgment

What it is

The document discloses that Claude Sonnet 4.5 retains residual safety risks, including susceptibility to adversarial prompting and jailbreak techniques, and that the hardcoded and softcoded safety behaviors may not be fully reliable in all adversarial contexts.

Why it matters (compliance & governance perspective)

This provision is a material disclosure of known limitations in the model's safety architecture, which is operationally significant for any operator making safety representations to their own users based on Claude's designed refusal behaviors.

Consumer impact (what this means for users)

Under this disclosure, users and operators should be aware that Claude Sonnet 4.5's safety behaviors, while structurally designed, are not claimed to be fully robust against all adversarial prompting techniques, and residual risk of safety behavior circumvention exists.

Institutional analysis (Compliance & governance intelligence)

1. REGULATORY LANDSCAPE: This provision engages with FTC guidance on substantiated safety claims and the EU AI Act's requirements for accuracy and robustness disclosures for general-purpose AI models. Where operators represent to their users that Claude categorically refuses certain content, the accuracy of that representation depends on the reliability of the hardcoded behaviors, which this provision qualifies. 2. GOVERNANCE EXPOSURE: High for operators who market their products on the basis of Claude's content safety. The disclosure that adversarial prompting may circumvent safety behaviors creates potential FTC exposure if operator-facing or consumer-facing safety representations overstate the model's actual reliability. Operators should review their marketing materials and user-facing disclosures for accuracy relative to this acknowledged limitation. 3. JURISDICTION FLAGS: EU operators subject to the EU AI Act's transparency and accuracy requirements should assess whether this residual risk disclosure affects their product conformity documentation. California operators should assess whether this disclosure triggers any obligation under California's AI transparency or consumer protection statutes. 4. CONTRACT AND VENDOR IMPLICATIONS: Operators should assess whether their commercial agreements with Anthropic adequately allocate liability for harms resulting from adversarial prompting that circumvents hardcoded safety behaviors. Standard API terms may not provide indemnification for such scenarios, and operators may face direct liability for harms to their own users. 5. COMPLIANCE CONSIDERATIONS: Compliance teams should review all user-facing and marketing representations about Claude's safety capabilities and assess whether they are consistent with this disclosed limitation. Organizations should establish incident response procedures for cases where adversarial prompting results in safety behavior failures in their deployments.

Applicable agencies

Provision details

Other risks in this policy

• Hardcoded Absolute Refusal Behaviors medium
• Operator and User Softcoded Permission Hierarchy medium
• Agentic Task Minimal Footprint Principle high
• Responsible Scaling Policy and Safety Level Thresholds medium
• Multi-Agent Trust and Prompt Injection Risk high
• CBRN and Cyberweapon Uplift Evaluation Disclosure medium