Hardcoded Absolute Refusal Behaviors

What it is

The document establishes a category of model behaviors that cannot be unlocked or modified by any operator or user instruction, including refusal to assist with weapons capable of mass casualties and refusal to generate sexual content involving minors.

Why it matters (compliance & governance perspective)

This provision establishes a non-negotiable safety floor across all deployment contexts, meaning operators building products on the Claude API cannot contractually or technically override these restrictions, and users cannot prompt the model past them regardless of framing.

Consumer impact (what this means for users)

Under this provision, users of any Claude Sonnet 4.5-powered product will encounter absolute refusals in defined categories regardless of the operator's configuration or the user's instructions, as the document states these behaviors are structurally fixed at the model level.

Institutional analysis (Compliance & governance intelligence)

1. REGULATORY LANDSCAPE: This provision engages with the EU AI Act's requirements for general-purpose AI models with systemic risk, which mandate that providers implement measures to prevent foreseeable misuse and assess systemic risks. The FTC's prohibition on unfair or deceptive practices is also relevant if safety claims about hardcoded restrictions are asserted to operators and users. The provision aligns with NIST AI RMF governance expectations for documented and tested safety controls. 2. GOVERNANCE EXPOSURE: Medium. The provision creates clear compliance value by establishing documented, tested refusal behaviors, but also creates exposure if adversarial prompting techniques demonstrably circumvent these restrictions in production, as the document itself acknowledges residual jailbreak risk. Organizations that represent to their own users that certain content is categorically unavailable are dependent on the accuracy of Anthropic's hardcoded behavior claims. 3. JURISDICTION FLAGS: EU operators deploying Claude in high-risk application categories under the EU AI Act may need to document and verify that hardcoded refusals function as described as part of conformity assessment. California operators subject to AB 2013 AI transparency requirements may need to reference this disclosure in their own AI transparency reports. 4. CONTRACT AND VENDOR IMPLICATIONS: Operators should assess whether their commercial agreements with Anthropic include representations and warranties regarding the performance of hardcoded refusal behaviors, and whether liability is allocated appropriately if a hardcoded refusal fails in a production deployment. Standard API terms may not provide sufficient indemnification for downstream harm resulting from a hardcoded behavior failure. 5. COMPLIANCE CONSIDERATIONS: Compliance teams should document their reliance on Anthropic's hardcoded behavior claims and establish independent testing protocols to verify refusal performance in their specific deployment context. Where products are marketed on the basis of content safety, legal teams should review whether safety representations are substantiated by the system card's disclosed evaluation methodology.

Applicable agencies

Provision details

Other risks in this policy

• 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
• Residual Safety Risks and Jailbreak Acknowledgment medium