Frontier AI oversight has three separate authorities: access, disclosure and enforcement

Britain's AI Security Institute can run sophisticated tests without possessing a general power to demand access to every frontier model or block its release. An outside evaluator may have laboratory access but limited rights to publish its results. Meanwhile, a privacy watchdog can investigate data use and a financial supervisor can finalize operating expectations for its regulated institutions. Each may be described as 'AI oversight', but they exercise materially different authority.

The October 13 UK parliamentary testimony, reported disputes over third-party evaluation contracts, the UK ICO's October 8 intervention and Singapore's October 7 supervisory guidelines provide enough evidence to separate access, disclosure and enforceable responsibility. Combining them as one new worldwide safety regime would obscure both existing legal duties and the gaps that remain.

The UK parliamentary committee asked what happens if a capable government evaluator identifies a serious risk but cannot stop a model's launch. According to WIRED's October 13 hearing account, Henry de Zoete of the AI Security Institute acknowledged that access depends on cooperation from developers and relevant foreign governments. Cat Little of the Cabinet Office confirmed that the current frontier-evaluation framework is voluntary. AISI has testing expertise, but it is not a general licensing authority with the power to refuse publication of a model.

The Anthropic Mythos 5.1 example makes that limitation specific. Liam Booth-Smith told lawmakers that an earlier US-government-related issue prevented full unrestricted UK AISI access and that access was subsequently granted. This establishes an earlier restriction and a reported change in status, not the legal basis for that restriction or the scope of today's testing rights. OpenAI's Tom Duff Gordon likewise discussed UK predeployment model access as subject to US-government requirements. Saying that the institute 'has access' does not by itself tell a reader whether it has all weights, tools, traces, dangerous-capability interfaces or sufficient time to run independent experiments.

The providers also disagreed about formal obligations. Anthropic supported binding requirements; OpenAI supported statutory elements including incident reporting. Google DeepMind emphasized safety standards, while Meta emphasized existing regulators. These positions are reported testimony, not a parliamentary enactment, and neither of the latter positions establishes opposition to every possible enforceable AI duty.

An independent lab may conduct a high-quality test yet face obligations limiting what it can publish. Reuters's October 13 reporting described negotiations over publication approval, proprietary details and evaluator access. OpenAI and Anthropic disputed interpretations suggesting an unrestricted veto. Complete signed agreements were not available in the reviewed material, and there is no supported finding that a model-safety result was falsified.

The distinction is between a narrow, reviewable redaction of a genuine trade secret and discretionary power to suppress an unfavorable conclusion. The first can preserve meaningful scientific scrutiny; the second weakens the public value of an evaluation even when private testing took place. The frontier evaluator dossier records the dispute without treating either side's contractual interpretation as proven. Publication independence also requires knowing when the assessor received tools and transcripts, whether the lab could rerun a finding and whether the final description may disclose material limitations.

The ICO's October announcement credits ten model developers with improvements already made or promised, while reporting separate ongoing inquiries into agent testing and deployment. The agency is asking whether the organizations assessed personal-data risks and safeguarded agent actions that reached external systems. It has not concluded that every lab contacted committed an unlawful act. A commitment to implement a control is also not identical to an implementation already in use. The ICO's evidence call runs to November 20, 2026 and is intended to inform later guidance and a statutory AI code.

Singapore's MAS guidelines are final rather than proposals, yet their consequences fall on a different group: financial institutions under MAS supervision. They must identify and assess AI risk, oversee third-party AI and apply lifecycle controls in proportion to materiality. Sections 3 and 4 should be met from October 7, 2027, with Sections 5 and 6 by October 7, 2028. MAS also plans more detailed consultation on agentic AI in 2027. These requirements do not depend on MAS independently discovering every flaw in a frontier model. They make the deploying institution responsible for the service it chooses to authorize, including where an external provider supplies the model.

When a government or developer announces 'independent AI safety testing', five questions locate the actual authority. Who can compel access? Is that right statutory, contractual or dependent on goodwill? What evidence is available? Are model behavior, tools, transcripts and incident traces within scope? Who controls disclosure? Can a reviewer report adverse outcomes without discretionary provider approval? Who has the remedial duty? Is it the model supplier, deployer or regulated institution? Who can impose consequences? Can a relevant agency require a change, investigate misuse or stop deployment, and under what existing law?

The UK hearing reveals a possible gap at the access and release-enforcement stages. The Reuters account raises an unresolved question about publication independence. The ICO process exercises data-protection scrutiny, while MAS places explicit AI risk-management responsibility on supervised institutions. The differences matter most for agents, because a model's benchmark score may be sound while the eventual deployment grants it operational credentials, external tools and authority over personal or financial data. A serious oversight claim has to specify which of those boundaries it can actually inspect and govern, instead of borrowing credibility from the phrase 'independent evaluation'.