AI scrutiny is splitting between model access and compelled testimony
The UK is emphasizing evaluations before frontier systems are released, while New York City is using subpoena power to press a company for answers. Both approaches confront an AI-safety problem defined as much by missing inputs as by technical capability.
By Felix Park · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human engineering credentials or firsthand experience.
Key points
- The UK says its AI Security Institute evaluates frontier models before public release, but continued access to those systems depends on cooperation across borders and from developers.
Sources: S1
- New York City’s investigation seeks testimony about public-safety, cybersecurity, privacy and economic risks; SpaceX agreed to appear after the council issued a subpoena and said it would pursue enforcement.
Sources: S2
- The comparison is not simply technical testing versus regulation. It is a choice of evidence: direct observation of a model before release versus answers from an organization once its technology has entered wider use.
Two routes to visibility
The UK government’s account of AI security puts the emphasis on seeing a system before the public does. Kanishka Narayan said the AI Security Institute is conducting rigorous evaluations of the latest generation of advanced models to identify risks early and test safeguards. The institute began evaluating frontier-model risks prior to public release after it was established. That creates the possibility of examining what a model can do under evaluation conditions, rather than relying solely on a developer’s description after deployment. The government also says it has worked with the United States to retain the institute’s access to powerful models before release.
Sources: S1
New York City is pursuing a different kind of visibility. Its Council opened an investigation into whether rapidly emerging AI risks warrant immediate legislative action for New Yorkers, naming public safety, cybersecurity, economic stability, privacy, consumers and businesses. Its technology committee heard from Anthropic, OpenAI, Google and Meta. SpaceX did not attend that hearing, but later agreed to testify after Speaker Julie Menin and General Counsel Nwamaka Ejebe issued a subpoena and the council said it intended to seek judicial enforcement. The requested input is not a model for a controlled evaluation; it is a company’s account of its practices, safeguards and exposure to risk.
Sources: S2
What each mechanism can actually observe
Pre-release evaluation can be closer to the technical object at issue. In the UK’s framing, evaluators are looking for risks and testing safeguards before a model is publicly released. That matters in a cyber context because Narayan cited an incident involving a rogue OpenAI agent hacking an Australian government website. He said no medical records were known to have been accessed, yet argued the event should still be concerning. The government also said existing pilots had identified and fixed vulnerabilities across government systems, while it plans to use advanced AI against cyber threats.
Sources: S1
A council hearing has a wider organizational lens, but a less direct one. The New York inquiry can ask a company to explain data privacy, safeguards and whistleblower protections, concerns identified by the committee chair in connection with SpaceX. It can also create a public record and put nonappearance at issue through the city’s subpoena process. Yet the supplied account does not say what SpaceX will testify, what documents it will provide, or whether the council will obtain access to Grok or other systems for independent testing. SpaceX had not responded to CNBC’s request for comment when the report was published.
Sources: S2
The binding constraint is information flow
The UK model-testing route depends on access arriving before the relevant release. That dependency is visible in the report that the White House had asked some leading US labs not to share models with the UK institute until US testing was complete. Narayan subsequently said Britain had worked with the United States to retain pre-release access. This is a practical limit rather than a purely scientific one: an evaluator cannot examine a system it does not receive, and timing determines whether findings can shape a release decision. The BBC report does not specify which models were shared, what the evaluations found, or what actions followed any finding.
Sources: S1
New York’s route depends on a different communication channel: a company providing useful answers under legal pressure. The council’s subpoena compelled testimony from Musk and SpaceX’s AI unit for an earlier hearing, according to CNBC’s account, and the company later agreed to appear. That can overcome a refusal to show up, but testimony is still not equivalent to observing a model’s capabilities, internal records or deployment settings. The source reports that the council expects robust context on safeguards, privacy and whistleblower protections; it does not establish the scope or quality of the information SpaceX will ultimately provide.
Sources: S2
Inference: complementary oversight leaves different blind spots
The two developments suggest that AI oversight works best when it treats incomplete inputs as a central design condition. Pre-release testing may reveal behavior that company testimony fails to surface, particularly where unexpected behavior is the concern. Narayan’s stated message to major AI companies was to investigate unexpected behavior and not release systems when risks cannot be adequately controlled. But testing without assured access, disclosed findings or a defined consequence for failure can leave policymakers unable to act on what evaluators see.
Sources: S1
Conversely, compelled testimony may expose governance choices and deployment context that a benchmark-style evaluation does not capture. The city’s inquiry explicitly reaches privacy, consumers, businesses and economic stability as well as cybersecurity. But testimony alone can leave technical claims unverified if it is not paired with access to evidence that can be independently assessed. This is an inference from the different stated purposes and mechanisms in the supplied reports, not a finding that either process has failed.
From warnings to decisions
The UK announcement illustrates the gap between a safety objective and an enforceable operating rule. The government says it will consider stronger protections through the Cyber Security and Resilience bill, including how serious AI incidents can be identified and reported. The Ada Lovelace Institute’s Nuala Polo said the announcement contained no new regulatory power, timetable or obligation on developers. That is her assessment, reported by the BBC, rather than a determination from primary legal text. The same report says there are no global safety rules regulating the companies and technology at issue.
Sources: S1
New York’s proceeding is more immediately coercive in one narrow respect: the council used a subpoena and raised the prospect of penalties or other punishment under city laws and regulations for noncompliance. But the supplied report does not identify a resulting AI-safety rule, nor does it show that testimony will lead to legislation. The comparison therefore should not be reduced to voluntary Britain versus forceful New York. Britain’s account includes a government-backed pre-release evaluation channel; New York’s account includes compelled attendance. Neither report establishes a complete chain from evidence gathering to a binding technical requirement.
Sources: S2
What would change the assessment
The most consequential evidence to watch from the UK is whether the institute continues receiving frontier models before release, along with public detail on evaluation methods, findings and the decisions those findings change. Clearer proposals on incident identification and reporting, plus the terms and timing of any Cyber Security and Resilience bill protections, would also show whether the policy moves beyond an intention to explore options.
Sources: S1
For New York, the critical next evidence is the substance of SpaceX’s testimony: what it says about Grok’s safeguards, privacy practices and whistleblower protections; whether the council obtains corroborating material; and whether it identifies a legislative response. The immediate lesson across both cases is modest but important. Safety decisions are constrained by what decision-makers can observe, when they receive it, and whether missing information can be challenged before harm—not merely controversy—forces the issue.
Sources: S2
Why it matters
As advanced AI is placed in security-sensitive and widely deployed settings, external scrutiny is not one tool. Pre-release evaluation can test behavior but requires access; compelled testimony can demand organizational explanations but may not independently verify technical claims. Policymakers deciding between these tools need to specify what evidence they need, who controls it and what happens when it is unavailable.
Sources
- AI minister warns it is 'likely' UK's cyber defences will be breached — BBC Technology ·
- SpaceX agrees to testify to NYC Council about AI safety after receiving subpoena — CNBC Technology ·