The Department for Science, Innovation and Technology ordered three frontier artificial intelligence labs to hand over United Kingdom red-team logs before expanding inference clusters beyond pilot capacity, according to letters seen by industry lawyers.
The notices, dated Wednesday, cover labs that already host evaluation hardware in Slough and Newport. Each must provide attack transcripts, refusal rates and compute-hour totals for models above ten-to-the-twenty-six floating-point operations before adding more GPUs.
Why logs matter
Ministers argue that voluntary safety tests mean little if scaling happens without regulators seeing failure modes. The AI Safety Institute will ingest logs directly rather than relying on corporate summaries.
One lab told staff the requirement mirrors banking stress-test data rooms, with strict confidentiality firewalls between commercial product teams and government reviewers.
Industry pushback
Trade association techUK warned that broad log sharing could expose trade secrets and deter investment in UK inference sites. DSIT responded that only safety-relevant prompts and outputs are in scope, not training corpora.
Timeline
Labs have until 21 November to comply or pause expansion plans. DSIT said non-compliance would trigger a public notice and could affect future grant eligibility under the AI Research Resource.
The opposition Liberal Democrats welcomed the move but asked whether Parliament will receive redacted versions for oversight.
Security testing detail
Red-team logs include prompt injection attempts, bioweapon synthesis probes and cyber-offensive coaching requests filtered by automated classifiers. DSIT wants raw transcripts rather than aggregated scores so the AI Safety Institute can replay attacks against updated model weights.
Lawyers for affected labs argued logs could contain customer data accidentally pasted into prompts; the department offered sanitisation guidance requiring personally identifiable information stripping before transfer, with penalties for non-compliance framed as breach of contract rather than criminal sanctions.
The policy intersects with the Data Protection Act because logs may include UK citizen inputs; DSIT said handling will follow restricted government-to-government channels similar to intelligence product sharing.
Start-ups without frontier-scale models are exempt, but any firm importing more than five hundred GPUs in a single shipment triggers notification duties under export control parallel rules announced in September.
Academic partners at Alan Turing Institute will receive redacted subsets for research on interpretability, though publication will require DSIT sign-off to avoid revealing attack surfaces.
Cloud providers hosting UK inference nodes said export control checks already delay GPU shipments by weeks; adding log audits could lengthen onboarding for new tenants. DSIT promised a four-hour SLA to review sanitized submissions so product launches are not frozen indefinitely.
European Commission officials watching the UK approach said Brussels may harmonise log-sharing rules in 2027, reducing patchwork compliance costs for labs operating in both jurisdictions.
Ministerial statements due next week are expected to frame the log requirement as proportionate to UK hosting benefits, including access to the AI Research Resource GPUs. Opposition Conservatives asked whether logs could be shared with allies under Five Eyes arrangements, a question DSIT declined to answer on the record.
Labs already publish voluntary safety cards; the government argues those summaries are insufficient when compute scale doubles within quarters rather than years.
Industry roundtables scheduled for Whitehall next Wednesday will let labs air implementation concerns before secondary legislation is drafted. Civil servants said criminal penalties are off the table for first-time log delays, focusing instead on expansion freezes.
Venture investors told founders that UK hosting clauses may become standard in term sheets, much like GDPR representatives became routine after 2018.
