OpenAI has revised its Preparedness Framework, the internal process it uses to judge AI model safety and decide what protections are required before and during release. The most closely watched change is a new opening for OpenAI to “adjust” its own safety requirements if a rival releases a “high-risk” AI system without comparable safeguards.
The update lands amid pressure on commercial AI developers to move quickly. OpenAI has faced accusations that it is lowering safety standards to speed up launches and that it has not delivered timely reports about safety testing. The company says it is not compromising on safety.
What OpenAI changed
The Preparedness Framework is OpenAI’s internal safety system for development and deployment decisions. In the updated version, OpenAI says the behavior of other frontier AI developers can affect how it applies its own safeguards.
The central point is not that OpenAI says it will automatically match a weaker rival. Instead, the company describes a conditional process. If another frontier AI developer releases a high-risk system without similar protections, OpenAI says it may change its requirements only after additional review.
“If another frontier AI developer releases a high-risk system without comparable safeguards, we may adjust our requirements,” wrote OpenAI in a blog post published Tuesday afternoon. “However, we would first rigorously confirm that the risk landscape has actually changed, publicly acknowledge that we are making an adjustment, assess tha t the adjustment does not meaningfully increase the overall risk of severe harm, and still keep safeguards at a level more protective.”
That language matters because it ties OpenAI’s safety posture to the wider market. The company is saying that its internal standards may respond to what competitors put into the world, but it is also saying that any such adjustment would need to be acknowledged publicly and judged against the risk of severe harm.
Why the rival-release clause matters
The update reflects a basic tension in advanced AI development: companies want to release models quickly, but they also need systems for evaluating harm before those models reach users. OpenAI’s new wording makes that tension explicit.
If a competing lab releases a high-risk system without comparable safeguards, OpenAI says the risk landscape may change. The logic is that a model already available from another developer could alter what protections are necessary or practical for OpenAI’s own systems.
At the same time, OpenAI says any adjustment would not be made lightly. The company lists several steps it would take first:
- Confirm that the risk landscape has actually changed.
- Publicly acknowledge that it is making an adjustment.
- Assess that the adjustment does not meaningfully increase the overall risk of severe harm.
- Keep safeguards at “a level more protective.”
Those conditions are meant to show that the framework is not simply a race-to-the-bottom mechanism. Still, the policy change gives OpenAI more flexibility than a fixed rulebook would, especially when competitors act first.
Automated evaluations take a larger role
The refreshed framework also signals that OpenAI is leaning more on automated evaluations. The company says it has not abandoned human-led testing, but it now points to “a growing suite of automated evaluations” designed to “keep up with [a] faster [release] cadence.”
That shift is important because testing speed is part of the broader release debate. Automated checks can be run repeatedly and quickly, which may help a company evaluate models at a faster pace. But the source article also notes reports that complicate OpenAI’s framing.
According to the Financial Times, OpenAI gave testers less than a week to complete safety checks for an upcoming major model, a shorter timeline than previous releases. The publication’s sources also alleged that many safety tests are now performed on earlier versions of models rather than the versions released publicly.
OpenAI has disputed the idea that it is compromising on safety. The company’s updated framework presents automation as a way to support faster development while maintaining safeguards, not as a replacement for all human review.
How OpenAI now describes model risk
The update also changes how OpenAI categorizes model capability and risk. The framework refers to models that may conceal their capabilities, evade safeguards, prevent shutdown, or self-replicate. OpenAI says it will focus on whether covered systems reach one of two thresholds: “high” capability or “critical” capability.
OpenAI defines a high-capability model as one that could “amplify existing pathways to severe harm.” A critical-capability model is one that could “introduce unprecedented new pathways to severe harm,” according to the company.
The distinction affects when safeguards are required. OpenAI says covered systems that reach high capability must have safeguards that sufficiently minimize the associated risk of severe harm before deployment. Systems that reach critical capability also require safeguards that sufficiently minimize associated risks during development.
In plain terms, the framework treats the most serious category as requiring protection earlier in the process, not only at the point of release. That is consistent with the idea that some AI risks may need to be managed while a model is still being built and tested.
The broader safety debate
The Preparedness Framework changes are OpenAI’s first since 2023. They arrive while the company is already under scrutiny over safety practices, release timelines, and corporate direction.
Last week, 12 former OpenAI employees filed a brief in Elon Musk’s case against OpenAI. They argued that the company would be encouraged to cut even more corners on safety if it completes its planned corporate restructuring.
Another criticism came from Steven Adler, who posted that OpenAI was quietly reducing its safety commitments and said one omitted change was no longer requiring safety tests of finetuned models.
Taken together, the update gives OpenAI a more adaptable framework for AI safeguards, especially when rival labs release high-risk systems first. The central question is how that flexibility will be used: as a disciplined response to a changed risk environment, or as another point of pressure in a market already moving fast.