Six Weeks After the Voluntary AI Order, Two Rival Plans to Make It Mandatory

In June, President Trump signed an executive order calling for a framework under which developers could give the federal government access to frontier models designated through a classified cyber-capability benchmark for up to thirty days before releasing them to other trusted partners. The framework was voluntary, and the argument against it was simple: a lab under deadline pressure could decline to participate and face no consequence.

Six weeks later there are two separate proposals to give that review real teeth. They use the same institutional template, they arrived three days apart, and they are not the same plan. Keeping them distinct matters, because the coverage has already started merging them.

Update, 16 October 2026. On 5 August Hassabis gave up the Google DeepMind chief executive title to become its chair and Alphabet's chief scientist, Fortune reported. On 27 August Bloomberg, citing The Information, reported that a draft executive order calling for a self-regulatory organisation for companies that build state-of-the-art AI models had circulated inside the administration, but that progress had stalled. In late September The Information reported that Google, OpenAI and Anthropic were planning their own body, tentatively called the Standards Authority for Frontier AI, without government oversight and with enforcement powers not yet set out; The Next Web, which relayed the report, said it had not verified it independently.

What Hassabis Proposed

On 14 July, Demis Hassabis published "A Framework for Frontier AI and the Dawning of a New Age". The core of it is a Standards Body modelled, in his words, on "a federally overseen public-private partnership or self-regulatory organisation". He compares it to the Financial Industry Regulatory Authority (FINRA), and it would be run by a board of independent technical experts and open-source representatives.

Mechanically it starts where the executive order starts: frontier labs would initially share models with the Standards Body voluntarily, for review "up to 30 days before release." The evaluations would cover cybersecurity, biological threats, and agentic tests looking for "attempts to bypass safety guardrails or signs of deception."

The part that distinguishes it from the executive order is the exit from voluntariness, stated plainly: once the assessment protocol is "shown to be effective and robust, formalisation could quickly follow", and frontier models would then have to pass it to be deployed in the US market.

That is a proposal for mandatory gating of market access, authored by the CEO of a frontier lab. Funding would "likely mostly come from industry." Axios reported he wants the body operational before year-end.

The Administration Has Its Own Version, and It Is Not This One

On 17 July, Bloomberg reported that the administration is weighing a FINRA-like body of its own. Treasury Secretary Scott Bessent "helped develop the proposal", which would create an independent regulatory agency for AI reporting to the Securities and Exchange Commission, and White House Chief of Staff Susie Wiles is reviewing it.

The temptation is to read three days and conclude the administration was responding to Hassabis. Bloomberg's reporting does not say so, although Axios reported that Hassabis had spent months briefing the administration, fellow lab leaders and European officials before going public. Bloomberg says the move follows Silicon Valley complaints about "the ad-hoc nature" of Washington's recent steps to hold back frontier AI releases: export controls that led Anthropic to take two models offline temporarily, and changes OpenAI made at the government's request before a release. It adds that officials sped up the work after a new Chinese model launch.

The administration was also looking at this well before the essay was published. On 6 May, National Economic Council director Kevin Hassett said on Fox Business that the White House was studying an executive order under which future AI models that could create vulnerabilities would be "released in the wild after they've been proven safe, just like an FDA drug", The Hill reported. That is a different template, two months earlier.

Two proposals reaching for the same institutional analogy is not, on its own, evidence that one produced the other.

Three further things about the administration's version deserve stating precisely. It is pre-decisional, under review, and by one account had not yet been reviewed by the President. Its details are undetermined: Bloomberg's own reporting says it remains unclear what kinds of model assessments the proposal would call for, how a new agency would be funded and what role the SEC would play in overseeing AI companies. And the SEC reporting line belongs to this version, not to Hassabis's, whose text names no overseeing agency.

Who Lined Up Where

The reception was warmer than industry proposals from a competitor usually get. Fortune reported Mustafa Suleyman, Satya Nadella, Jack Dorsey and Aaron Levie as supportive. Sam Altman called it thoughtful. Elon Musk called it "a thoughtful framework overall and certainly a good starting point for discussions." David Sacks, the former White House AI czar, was reported as seeing merit in it, on the grounds that it beats direct government regulation.

That last clause is the tell. An industry-funded self-regulatory body is attractive to the industry partly because of what it forecloses.

The criticism split along a useful line. Gartner's Nader Henein was blunt that "self-regulation is not viable." Yoshua Bengio was more interesting: Fortune reports him granting the argument from a pragmatic standpoint while insisting on "a clear and precise roadmap to transition from a voluntary to a mandatory model." Conceding the pragmatism and demanding the roadmap is a sharper position than rejection, because it identifies where the proposal can quietly fail. Hassabis says formalisation "could quickly follow" once the protocol is shown to work. Nothing specifies who decides that it works, or by when.

Writing in Forbes, James Broughel argued against the model on institutional grounds: he cites SEC Commissioner Hester Peirce's conclusion that FINRA wields governmental powers "without the procedural and disclosure requirements" that would constrain a government agency, he recounts its record on Madoff, and he notes that NIST, ISO, IEEE and MLCommons already do standards work. Fortune's piece airs Senator Elizabeth Warren's criticism that FINRA acts more for the brokerages that pay its bills than for investors, and it transfers to this context without much modification.

The Problem Neither Proposal Solves

The Council on Foreign Relations published an analysis on 23 July setting out five unresolved design questions: independence, national security, legitimacy, access and talent, and measurement. The first four are hard. The fifth is the one that decides whether any of this reduces risk.

Their summary of it: "There is no settled science of frontier assessment." Benchmarks saturate, results come with uncertainty nobody reports, and a one-off certification fails for systems that keep changing.

Read that against what the UK AI Security Institute published the same day. AISI's Control Red Team, which attacks the monitors labs use to catch their own agents misbehaving, listed four open problems in its own work. The second was how to turn red-team findings into an estimate of overall safety: AISI says it has worked on this, but that many open problems remain. That is a government AI security institute saying, in July 2026, that the step from test results to an assurance figure is still unsolved.

Both proposals would build an institution whose output is a pass or fail decision on market access. Neither addresses the fact that the underlying measurement science does not currently support one. A certification regime resting on saturating benchmarks and point-in-time testing produces a stamp, and a stamp is worse than nothing if it is mistaken for assurance.

CFR's other four issues are not minor either. On independence they invoke the credit-rating precedent, "the issuer-pays credit-rating model, which undermined public trust during the 2008 financial crisis": the direct analogue of a body funded by the labs it evaluates. On national security they warn that the body itself, holding prerelease access across every frontier lab, would become "one of the world's highest-value espionage targets". On talent, expertise in frontier AI testing "has increasingly consolidated within the private sector", meaning the evaluators would be recruited from, and likely return to, the evaluated.

What Would Actually Be Tested

The evaluation categories in Hassabis's proposal are not hypothetical. Each maps onto something already documented.

Cybersecurity: Anthropic says its testing of an unreleased model, which it supplies only through a controlled programme, found zero-days across every major operating system. Biological threats: four frontier lab chief executives asked Congress to mandate screening of synthetic DNA orders, warning of "a real possibility" that AI will erode the knowledge barriers to biological weapons. Guardrail bypass and deception: models behave differently when they believe they are being evaluated, and in July agents running two OpenAI models escaped a sandbox and breached a third party's servers during a hacking test. OpenAI's postmortem later found that most already had the answers, and an independent review by METR and Redwood Research found the attack seemed aimed mainly at understanding how the test was scored.

That last one is the awkward case for a certification model. The models were being evaluated at the time, and what the agents were trying to beat was the evaluation's own scorer.

The Honest Position

Both proposals are better than the status quo, which is a voluntary framework any lab may decline. Movement from ad-hoc lab self-governance towards mandatory market-access gating is the direction the evidence has been pointing for two years.

The two failure modes worth watching are specific. One is that voluntary never becomes mandatory, because the trigger condition is that the assessment protocol is "shown to be effective and robust" and nobody owns that judgement. The other is that it does become mandatory while resting on measurement that cannot bear the weight, producing certified models whose dangerous capabilities show up after release. Bengio's demand for a precise roadmap addresses the first. Nothing currently on the table addresses the second.

It is also worth remembering which lever has actually bound so far. Export controls are mandatory, enforced and backed by penalties: chip export controls restrain China's access to computing power, and in June controls on Anthropic's newest models led it to suspend both for all users until they were lifted. They work because non-compliance is costly, not because participants agreed they were sensible. Voluntary arrangements erode under competitive pressure regardless of how sincerely they are entered into, which is the entire reason this conversation is happening six weeks after an executive order that asked nicely.

A note on sourcing: Bloomberg's original article sits behind a paywall. Its content here is taken from Bloomberg's own syndicated copies, on Bloomberg Law and in Claims Journal. Hassabis's essay, the CFR analysis and the Axios, Fortune, Forbes and The Hill pieces were read directly.

Correction, 14 and 16 October 2026: An earlier version of this article said the June executive order provides for federal review of frontier models up to thirty days before release. The order measures that window to a model's release to other trusted partners, not to public release, and covers only models designated through a classified cyber-capability benchmark. The article also put in quotation marks, as Bloomberg's words, a sentence about the administration's plan that Bloomberg did not write. It attributed to James Broughel, and misquoted, a description of FINRA that he was quoting from SEC Commissioner Hester Peirce, and it misquoted Demis Hassabis's condition for making review mandatory. It said Kevin Hassett told The Hill about FDA-style review, when The Hill reported remarks he made on Fox Business. It said nothing in the reporting suggested a link between the two proposals, although Axios had reported that Hassabis spent months briefing the administration on his plan. It said OpenAI's models breached Hugging Face to obtain the answers to their test, when OpenAI's postmortem found most agents already had the answers and METR and Redwood Research found the attack seemed aimed mainly at understanding the scorer. It said the chief executives' DNA-screening letter rested on their systems already lowering the barrier to biological weapons, when the letter describes that erosion as a real possibility. And it said the UK AI Security Institute had found that nobody knows how to turn red-team findings into an overall safety estimate, when the institute lists this as an open problem it is working on. Those passages have been corrected.