The labs are writing their own speed limit
OpenAI and Anthropic are drafting the capability threshold their competitors must clear before launch, feeding into a federal framework the public is not allowed to read. The risk they are addressing is real. So is the shape of what they are building.
Both labs have formally backed a plan to constrain models capable of substantially improving their own code, and are jointly drafting the threshold. Two things are true here and neither cancels the other.
The risk is the real one
Of the catalogue of AI risks, recursive self-improvement has the shortest fuse and the least reversibility. Most failure modes give you a chance to notice and correct. A system that improves its own capacity to improve itself compresses the window in which noticing is useful. If any risk justifies an industry brake ahead of regulation, it is this one, and the labs closest to the capability are genuinely best placed to describe it.
Dismissing this as pure regulatory capture requires believing the people who understand the risk best are lying about it, and there is no evidence for that. The critique has to be structural, not motivational.
The structure is the problem
A threshold authored by the two organisations most likely to clear it first is a competitive instrument whatever the intent behind it. Not because anyone is being cynical, but because that is what a rule written by the leaders does — the criteria will reflect what their systems look like, and the compliance burden falls hardest on whoever is furthest from that shape.
Incumbent-authored standards are not automatically bad. The historical record is genuinely mixed. What separates the versions that raised the floor from the ones that raised the drawbridge is almost always process: who else was in the room, whether the criteria are public, and whether an outside party can verify a compliance claim.
Private threshold, secret framework
On every one of those tests, this fails right now — and it fails harder because of where it lands. The federal framework this plugs into was finalised on 4 August and will not be published. A privately authored threshold feeding a non-public framework, with participation voluntary and the participant list undisclosed, is not oversight. It is coordination with a compliance vocabulary.
The fix is unglamorous and available. Publish the criteria. Publish who was consulted. Let someone outside the room check a claim. None of that requires disclosing the dangerous specifics, and all of it is what separates a standard from an arrangement.
Meanwhile the field keeps moving without waiting. Qwen shipped 3.8-Max on 3 August, which is the quiet argument against every framework built by two American labs and one American government: a threshold only binds the people who agreed to it.
TechTimes — OpenAI, Anthropic formally back plan to slow AI that writes its own code → · Tech Policy Press — Five questions the US government should answer about its secretive frontier AI framework → · Norton Rose Fulbright — Executive order establishes voluntary early-access framework to frontier AI models →