Why These Frameworks Matter to Enterprise Buyers
Enterprises procuring frontier models from Anthropic or OpenAI are increasingly asked to evaluate vendor safety posture as part of AI risk assessment. Both companies publish internal governance frameworks describing how they classify model risk and what actions those classifications trigger before a model is trained further or deployed. For governance teams, these documents are useful inputs, but they are not compliance certifications, contractual guarantees, or substitutes for enterprise-side runtime governance. Understanding the structural differences between the two frameworks helps risk teams ask better vendor questions and identify where their own governance responsibility begins.
How the Risk Models Actually Differ
Anthropic's approach ties safeguards to a broad security and deployment posture. An ASL tier reflects not just what a model can do, but what security measures (such as model weight protection) and deployment restrictions must be in place. OpenAI's approach is more granular at the category level: each tracked risk category receives its own score, and the framework gates decisions based on the highest post-mitigation score across categories rather than a single overall tier. This means a model could be rated differently across Anthropic's and OpenAI's systems in ways that do not map one-to-one. Governance teams comparing vendors should treat ASL levels and category scores as different measurement systems, not equivalent labels, and build an internal translation layer when conducting comparative vendor risk assessment rather than assuming parity.
| Dimension | Anthropic RSP | OpenAI Preparedness Framework |
|---|---|---|
| Measurement unit | Single overall ASL tier (ASL-2, ASL-3) | Independent score per risk category |
| Gating logic | Tier reflects capability plus required security posture | Gated by highest post-mitigation score across categories |
| Cross-vendor comparability | Not directly equivalent to category scores | Not directly equivalent to ASL tiers |
What Changed in the Past 12 Months
Both frameworks have been revised within the past year. Anthropic updated its RSP alongside the activation of ASL-3 protections for specific Claude model releases, reflecting changes to deployment and security standards at that tier. OpenAI published a Version 2 update to its Preparedness Framework, including changes to the Safety Advisory Group's role and clarified escalation paths between technical safety teams and company leadership. Neither company has published a complete version-diff explaining exactly what changed relative to prior versions, and neither framework has documented quantitative thresholds for all risk categories. Enterprises relying on these frameworks for vendor risk assessment should track framework version history over time rather than treating a single point-in-time review as sufficient, since both companies retain unilateral authority to revise thresholds and categories.
Where Vendor Frameworks Stop and Enterprise Governance Begins
Both the RSP and the Preparedness Framework focus on catastrophic, frontier-level risks such as CBRN misuse, cyber capability, and autonomous replication. Neither addresses enterprise-relevant operational risks such as data leakage, prompt injection, tool misuse, or business-logic errors introduced by an agent's specific configuration. These frameworks also govern the base model at the provider level; they do not extend to how an enterprise configures, permissions, or monitors an agent built on top of that model. A model classified under ASL-2 or scored low across OpenAI's categories may still be deployed by an enterprise with excessive tool permissions, no audit trail, or no runtime monitoring of agent behavior. The safety classification describes the model as evaluated by the vendor; it does not describe the risk profile of the enterprise's specific deployment.
The Practical Governance Gap
The core takeaway for enterprise AI governance leaders is that provider-level safety frameworks and enterprise-level runtime governance solve different problems. Anthropic's RSP and OpenAI's Preparedness Framework are designed to prevent the training or release of models that cross catastrophic capability thresholds. They were not designed to give enterprises visibility into agent behavior at runtime, enforce least-privilege access for specific tools, or produce an audit trail for a regulator or internal compliance review. This is the gap enterprises must close independently: identity and permissions for each agent, approval workflows for sensitive tool use, monitoring of agent-to-agent and agent-to-tool interactions, and logging sufficient to reconstruct what an agent did and why. Trussed AI provides runtime governance and security for enterprise AI agents, including permissions management, tool approval workflows, and audit logging, addressing the operational layer that sits below vendor-level safety frameworks regardless of whether the underlying model comes from Anthropic, OpenAI, or another provider.