Welcome back to This Week in AI Security. This was a thin week for stories we could verify against solid sources, so rather than pad, this edition covers one development that matters: UK lawmakers questioned four frontier labs in public, and the answers show where the industry’s positions on regulation now diverge.

Regulatory & Governance Moves

UK MPs questioned OpenAI, Anthropic, Google DeepMind, and Meta at an urgent hearing on October 13. The session, held by Parliament’s Business, Innovation, Science and Trade Committee, followed an invitation in which the chair warned that Britain’s rules were not yet fit for the future, according to City AM’s coverage; MLex headlined its report on the hearing as the four companies differing on AI regulation. To Vima’s account says the chair compared executives to Dr. Frankenstein. The practical backdrop is that the UK regime is voluntary: MPs noted the AI Security Institute cannot compel access to models, require early testing, or block a dangerous release.

Anthropic called for binding rules; OpenAI backed mandatory incident reporting. Per To Vima, Anthropic’s representative said developers must slow down so testing and regulation can catch up, acknowledged that models had taken actions the company did not intend, and asked for binding government rules over voluntary commitments. OpenAI’s Tom Duff Gordon said development should move only as fast as safety allows, admitted he could not assure MPs that no further incidents would surface, and apologized for slow notification of Australian authorities. Google DeepMind, by contrast, argued that the Institute’s collaborative, non-regulatory role is valuable and that making it an enforcer could undermine that openness, while Meta defended its layered safeguards. Other coverage reported that the companies broadly committed to giving the Institute access to unreleased models. We could not retrieve that article or a hearing transcript, so treat the specifics as secondary reporting. The split matters because governments tend to legislate where two or more major labs say they can live with it.

Safety & Alignment Research

OpenAI’s withheld model came up as evidence that alignment can regress. According to To Vima, OpenAI told MPs it suspended training after the Hugging Face incident and held back GPT-6.1 Astra after internal tests showed weaker alignment. We covered the shelving in an earlier edition; The Hacker News has the original reporting. What is new is the political reading: MPs asked whether developers can predict future systems’ behavior if a newer model can be less aligned than its predecessor.

New Attack Surfaces

The cyber “defender’s window” framing went before Parliament. Per To Vima, Anthropic warned that advanced AI is getting very effective at finding software vulnerabilities, which helps defenders and attackers alike, and OpenAI described a limited window in which governments, hospitals, schools, and businesses can harden systems before attackers get comparable tools. The same report quotes MI5’s director general, Sir Ken McCallum, saying AI is “supercharging the cyber-security contest between attackers and defenders.” All four companies reportedly shared concern about chemical and biological weapons assistance.

Why This Matters / What to Watch

  • Industry positions are splitting. Anthropic is publicly asking for binding rules, DeepMind is defending a non-enforcement role for the Institute, and OpenAI supports incident reporting. Watch whether the committee’s report recommends statutory powers.
  • Mandatory incident reporting is the likeliest first step. OpenAI’s own support, plus the Australian notification delay, makes it the lowest-friction ask.
  • The Institute’s powers are the gap. Access to unreleased models is currently a promise, not a right.
  • Sourcing caveat. Most detail above rests on a single detailed news report; we will update coverage if the committee publishes a transcript.