OpenAI rates new models as 'High risk' for cyber and bio skills but not for self-improvement — what actually earns that label?

This explores what OpenAI actually measures when it rates a model's ability to improve AI systems, including itself, and whether that measure captures what researchers mean by self-improvement. The short answer is that the collection doesn't contain the framework's formal definition. It does show how the framework is applied in practice, and that is the more revealing part. In its assessment of GPT-5.6 Sol and Terra, OpenAI rated both models High in cybersecurity and biorisk but below High in AI self-improvement, even though both did measurably better on internal research-debugging tasks Does GPT-5.6 show meaningful self-improvement capability?. The surprising detail is that the self-improvement rating rests on one debugging metric, and the published material doesn't put numbers on it. So in practice, "self-improvement capability" is measured by whether a model can help with the routine work of AI research. It is not measured by whether the model can redesign itself.

Why measure debugging instead of something more dramatic? One reason is that real self-improvement probably doesn't start with a model rewriting its own weights. Lilian Weng argues that the near-term path runs through the "harness": first the prompts, then the code that wraps and deploys a model, then the optimizer code Does recursive self-improvement start with harness engineering?. Concrete systems already work this way. The Darwin Gödel Machine improves coding agents by testing variants against benchmarks and keeping an evolving archive of the versions that work, roughly doubling performance on SWE-bench Can AI systems improve themselves through trial and error?. A "bilevel" setup goes a step further: an outer loop reads the inner loop's code, finds the bottlenecks, and writes new search methods while it runs Can an AI system improve its own search methods automatically?. A debugging metric looks for the early, mundane signs of this kind of capability.

The catch is that "self-improvement" covers very different things. A survey of 1,250 papers separates bounded self-refinement from open-ended recursive self-improvement Are self-refinement and recursive self-improvement actually the same thing?. Bounded self-refinement means a model polishing its outputs against a checkable target, which is common industry practice today. In open-ended recursive self-improvement, each round of improvement makes the next round more capable. The survey argues that grounding requirements, collapse dynamics and compute limits still hold that second kind back. Other researchers suggest the real turning point is something a debugging task can't detect: whether the AI chooses its own research goals without drifting Can AIs learn to specify their own research objectives?. A related view is that a system truly improves itself only when it can change how it learns. Today's systems run human-designed reflection loops that break when the domain shifts Can AI systems improve their own learning strategies?. Work on agents given only vague goals points the same way. Much of the hard work is turning "get better at X" into something measurable, and that step comes before any optimizing Can agents learn from vague goals without predefined metrics?.

This matters because a narrow benchmark can mislead in either direction. Neat, auto-gradable tasks can overstate or understate what frontier models can do. Open-ended evaluations of messy, long-horizon work tend to spot new capabilities earlier Do automated benchmarks hide what frontier AI systems can really do?. A single debugging score may be exactly the kind of instrument that misses a capability as it first appears. The stakes are not abstract. Anthropic has warned publicly that recursive self-improvement carries serious societal risks and has suggested slowing some development paths Does recursive self-improvement pose serious risks to society?. The Future of Life Institute argues that companies can't police this alone and should face government limits Can companies alone manage the risks of AI systems?. The takeaway is that the "below High" rating answers a narrow question: can this model debug AI research code at a certain level? The bigger questions, such as whether a model can set its own goals, change how it learns, or improve the scaffolding around it, are being studied in the research literature. The collection shows no sign that they are part of the framework's threshold.