Public Policy & Governance

The US government issued an export control directive to Anthropic ordering suspension of access to its most advanced models, Fable 5 and Mythos 5, for all foreign nationals. Anthropic stated it had to 'abruptly disable' the models for all customers to ensure compliance, as it lacked the technical infrastructure to enforce nationality-based access controls at scale. The government's stated rationale — that safeguards can be bypassed and the models used to identify software vulnerabilities — signals that national security agencies are now treating frontier AI capabilities as equivalent to controlled dual-use technologies. The Guardian and Politico both confirm the directive was issued without specific details being provided to Anthropic about the underlying intelligence.

This is the first confirmed instance of a commercial AI lab being ordered to restrict model access via export control mechanisms rather than voluntary policy. The enforcement mechanism — disabling access for all users rather than just foreign nationals — exposes a critical implementation gap: AI labs currently lack the identity-verification infrastructure to comply with nationality-based restrictions without blunt, market-wide shutdowns. This will almost certainly force investment in access-control architectures across the industry, and sets a template that other agencies — and allied governments — will observe closely.

Anthropic CEO Dario Amodei publicly endorsed mandatory safety vetting for frontier AI models, a position that places Anthropic in direct contrast with competitors who have resisted binding pre-deployment requirements. Amodei also proposed 'universal capital accounts' as a fiscal policy response to AI-driven job displacement. Politico reports the statements were made in a policy-facing context, suggesting deliberate stakeholder positioning rather than casual remarks.

Anthropic's endorsement of mandatory testing carries strategic logic: as a lab that has built its brand on safety, it benefits from regulatory requirements that incumbents with established safety infrastructure can meet more easily than new entrants. This is a textbook regulatory capture dynamic and should be read as such by policymakers. However, it also provides political cover for legislators who want to move on mandatory evaluation frameworks — having a major lab on record in support removes one line of industry opposition.

The White House held separate meetings this week with children's advocates and tech industry representatives as part of a coordinated effort to build support for federal preemption of state AI laws. Politico reports this came days after a bipartisan House proposal on AI received a cold reception, suggesting the administration is seeking a narrower, more politically durable hook. Children's safety has served as the entry point for federal digital regulation before — the Children's Online Privacy Protection Act and subsequent attempts to update it followed the same pattern.

The strategic logic is clear: framing preemption around child safety splits civil society opposition, as groups focused on child protection may accept federal standards in exchange for a single enforceable baseline rather than a patchwork. The risk for states with more ambitious AI laws — California's SB 1047 framework being the most notable — is that a federal floor set low enough to satisfy industry becomes a ceiling that displaces stronger state protections. This tension between floor and ceiling is the central contested question in any preemption debate.

The European Commission published the final Code of Practice on marking and labelling of AI-generated content, with the AI Office organizing a signature session on June 22 for potential signatories. European Commission confirms the Code is voluntary and sets out practical implementation steps. This is a deliverable under the AI Act's broader transparency obligations and follows the Commission's established pattern of using voluntary codes to operationalize Act requirements before mandatory enforcement kicks in.

The voluntary nature is the critical caveat. The Code sets industry norms and gives the Commission evidence of what is technically feasible — information it will use when drafting binding implementing acts. Companies that sign and meet the Code's standards reduce their regulatory risk; those that don't will face scrutiny when mandatory labelling requirements take effect. The June 22 session is therefore as much a political event as a technical one — the list of signatories will signal which platforms are engaging constructively and which are not.

A Derbyshire police officer is under criminal investigation for allegedly using AI to create fabricated evidential material across multiple cases and perverting the course of justice — the first known criminal investigation of this type in the UK. The Guardian reports the officer has been removed from frontline duties. Separately, a Florida man is suing law enforcement agencies after a facial recognition algorithm incorrectly matched him with a 93% confidence score to a suspect in a crime committed 300 miles away, leading to wrongful arrest and prosecution. The Guardian covers the Florida case.

These two cases represent distinct but converging governance failures. The UK case is a malicious use by a public official — a problem of procurement, access controls, audit trails, and internal oversight. The Florida case is a systemic reliability failure — a problem of procurement standards, validation requirements, and judicial gatekeeping of AI evidence. Neither jurisdiction has mandatory pre-deployment standards for law enforcement AI that would have caught either issue. The UK's absence of a binding police AI framework and the US's fragmented, jurisdiction-by-jurisdiction approach to facial recognition regulation are directly implicated.