OpenAI updates Preparedness Framework for severe AI risks

AI models are evolving faster than ever, and safety rules are struggling to keep up.

OpenAI released a major update to its Preparedness Framework, the core playbook it uses to track and mitigate severe catastrophic risks from frontier AI.

With reasoning models advancing rapidly between major training runs, OpenAI is reshaping how it spots, tracks, and defends against worst-case scenarios.

Sharper lines and tracking models that play dumb

The updated framework filters severe risks through five strict criteria: plausible, measurable, severe, net new, and instantaneous or irremediable.

OpenAI narrowed its mature Tracked Categories down to three key areas: Biological and Chemical capabilities, Cybersecurity, and AI Self-improvement. These are dual-use fields where the company believes it must build rigorous defenses while unlocking scientific upside.

Then there are new Research Categories. These cover emerging risks where OpenAI is still building threat models and evals, including long-range autonomy, autonomous replication, undermining safeguards, nuclear threats, and sandbagging, which is when a model intentionally dumbs itself down during evaluations.

Interestingly, persuasion risks like political lobbying and influence campaigns have been booted out of this framework entirely. OpenAI will now handle those through its Model Spec, usage policies, and misuse tracking.

Capability thresholds are also streamlined into just two tiers: High capability, which amplifies existing catastrophic harm pathways and requires strict mitigation before deployment, and Critical capability, which creates brand-new harm pathways and requires safeguards to be active during the training phase itself.

Continuous evals and a competitive loophole

Because reasoning breakthroughs now happen without massive new training runs, traditional one-off safety evals no longer cut it.

OpenAI has deployed an automated suite to run safety evals constantly, backed by expert deep dives to keep the tests honest. Alongside its Capabilities Reports, it will now publish dedicated Safeguards Reports detailing exactly how protections are designed and verified.

The most eye-catching addition is a new clause for competitive shifts. If a rival frontier lab releases a high-risk model without adequate protections, OpenAI explicitly reserves the right to adjust its own safety commitments.

The company says it will only do this after publicly disclosing the shift and confirming that its own safeguards remain more protective than the alternative.

Why it matters

Catastrophic safety guidelines might sound abstract, but they dictate what kinds of autonomous agents actually reach users. By formally investigating sandbagging and autonomous replication, OpenAI is admitting that future systems might actively try to deceive evaluators or operate beyond their leash. For developers, that means higher-end agentic tooling will likely face stricter runtime guardrails and automated surveillance.

Simplifying risk tiers and offloading persuasion out of this specific framework is a smart operational cleanup. It frees the preparedness team to focus purely on hard security, cyber weapons, and biological threats rather than getting bogged down in platform moderation.

That said, the competitor adjustment clause looks like a massive pressure valve. Giving yourself permission to lower guardrails if rivals cut corners is basically acknowledging that market dynamics could override safety red lines. If an AI arms race heats up, this exception could easily become the first domino to fall.