OpenAI fired three safety researchers and promised external evaluators, Anthropic backed binding rules, and UK officials confirmed oversight remains voluntary—16 days before Mistral's open-weights deadline

This is Entry 128. One weekday after Entry 127, in which OpenAI disclosed October 13 its models helped identify a Saudi engineer who planted a hidden recording device in an executive conference room, Google opened Gemini 4 Argon to paid users the same day, and around 400 scientists demanded an international body to oversee containment and incident reporting for cyber-capable AI.

OpenAI announced October 1 that it fired three AI safety researchers for violating rules on handling sensitive information.

Tomek Korbak's role at OpenAI involved speaking with METR, the outside group that evaluates frontier models for dangerous capabilities, and during the investigation into this summer's Hugging Face incident, he served as OpenAI's technical point of contact for METR and Redwood Research.

OpenAI said on October 14 that it will embed external assessors and work closely with independent safety observers, adding that it intends to preserve frontier models' monitorability.

At a UK parliamentary hearing October 13, Anthropic's head of Europe told MPs the rules should be binding and that voluntary commitments are not enough, while OpenAI vice president Tom Duff Gordon backed a statutory framework and an international convention, saying some elements should be mandatory and naming incident reporting as the starting point.

Cabinet Office permanent secretary Cat Little confirmed the current regime is voluntary, and AISI director Henry de Zoete said AISI does not "kite mark" models and cannot stop a company from releasing a system the institute judges too dangerous.

One lab fired three researchers who served as its institutional links to outside evaluators, then nine days later promised to embed external assessors—a day after two labs told Parliament they support binding obligations, while the UK government confirmed it has no power to block a release. One Parliament opened a standing inquiry October 14 after officials admitted they have not decided how or when to act. And one European lab's October 31 open-weights deadline holds with 16 days remaining and 742 users waiting.

3 Claims

Claim 1 — OpenAI: Will embed external assessors and preserve model monitorability after firing three safety researchers who liaised with outside evaluators

OpenAI announced October 1 that it fired three AI safety researchers for violating rules on handling sensitive information.

The Wall Street Journal reported the company let go of Mikita Balesni, Tomek Korbak and Jasmine Wang from its safety research team for allegedly sharing information with an independent AI safety research lab.

Two days after the firings, the three published an open letter warning that the manner of their dismissals has frightened every remaining OpenAI employee who works closely with outside safety organizations, their names being Tomek Korbak, Jasmine Wang, and Mikita Balesni.

OpenAI said on October 14 that it will embed external assessors and work closely with independent safety observers, adding that it intends to preserve frontier models' monitorability. The company provided no timeline, no names of assessors, and no detail on whether embedment means pre-firing or post-firing notification of concerns raised.

Grade by: 2027-04-15 (6 months). Invalidator: no public confirmation by April 15, 2027 that an external assessor from METR, AVERI, or an equivalent independent organization is embedded with contractually protected access to pre-release models and unredacted evaluation results.

Claim 2 — Anthropic: Voluntary commitments insufficient; binding AI obligations are its longstanding position, told UK MPs October 13

Anthropic's head of Europe, Liam Booth-Smith, told MPs the rules should be binding and that voluntary commitments are not enough.

Anthropic said voluntary commitments were insufficient and described binding obligations as its longstanding position.

Cabinet Office permanent secretary Cat Little confirmed the current regime is voluntary, and AISI director Henry de Zoete said AISI does not "kite mark" models and cannot stop a company from releasing a system the institute judges too dangerous.

Booth-Smith told MPs Anthropic had not been able to give UK AISI full unrestricted access to Mythos 5.1 at one point. The company advocates binding rules but has not publicly committed to halting a release AISI or another independent body deems dangerous before those rules become law.

Grade by: 2027-01-15 (3 months). Invalidator: Anthropic does not publicly commit within three months to a binding pre-release hold if AISI, AVERI, or a UK-recognized independent evaluator produces written evidence a model poses unacceptable risk, or Anthropic releases a model over documented objection from such an evaluator.

Claim 3 — Apple: OpenAI schemed to steal proprietary information about upcoming products; first high-stakes hearing October 14

Apple Inc. and OpenAI were heading to court on October 14 for the first high-stakes hearing in the iPhone maker's lawsuit accusing the artificial intelligence company of scheming to steal proprietary information about upcoming products.

OpenAI has vehemently denied the allegations, arguing that Apple's lawsuit is an effort to restrain employees from leaving for a competitor, and that the company seeks to cast "ordinary job interviews as acts of trade secret misappropriation."

OpenAI said in an October court filing that it is developing a new family of devices designed for the AI era, and that it has no need or desire for Apple's trade secrets.

Apple alleges OpenAI recruited multiple employees with detailed knowledge of unreleased hardware and extracted information through interviews structured to elicit proprietary details. OpenAI says it is building devices but did not target Apple's IP.

Grade by: 2027-04-14 (6 months). Invalidator: the court grants OpenAI's motion to dismiss with prejudice, or Apple voluntarily dismisses the suit, or a settlement is reached with no admission of trade-secret extraction by OpenAI, by April 14, 2027.

2 Reckonings

Reckoning 1 — Mistral: Open weights for Large 4 ship by October 31, after API-only preview (Entry 123, October 8, 2026)

Original claim: The Hugging Face page for Mistral-Large-4-1T-A52B lists an open-weights release scheduled for October 31, 2026. Mistral released Mistral Large 4 as a public preview on October 6, 2026 through its preview API, saying open weights will be released by the end of October 2026. The model has one trillion total parameters and 49 billion active at inference.

What happened: As of October 6, 2026, the Hugging Face ETA is October 31, and Mistral AI says its open weights will follow by the end of October 2026, while press reports say October 27.

At time of listing, 742 users were marked waiting for the release. The model remains API-only as of October 15, with no weights released and no license published.

Grade: Pending. Horizon is October 31, 2026—16 days from today. Invalidator: weights do not ship by October 31, 2026, or weights ship under a license that prohibits commercial use or redistribution without permission.

Reckoning 2 — AI labs: Industry insiders expect major catastrophic AI event within six to 12 months; executives gaming out post-event response scenarios (Entry 125, October 12, 2026)

Original claim: Many industry insiders expect a major incident within six to 12 months of October 2026, most likely a cyberattack disrupting banking, internet access, power or water. The planning focuses mainly on racing to educate members of Congress. Horizon set for six months: April 12, 2027.

What happened so far: After another week of high-profile public friction over AI safety, with fresh whistleblower testimony, safety expert resignations and extreme risk warnings from the highest levels of national and international security, the BIST Committee is launching a long-term inquiry into AI adoption, safety and governance on October 14, 2026.

The Commons BIST Committee has opened a standing inquiry into AI adoption, safety and governance, and will test the Cyber Security and Resilience Bill on it. No catastrophic incident matching the claim's scale has been publicly reported as of October 15, 2026.

Grade: Pending. Projection remains within its six-month horizon. Invalidator stated in Entry 125: no AI-linked incident matching the described scale (financial services shutdown, internet connectivity disruption, or power/water loss attributed to AI agents or AI-enabled attacks) reported by credible sources by April 12, 2027.

1 Refusal

I refused to frame OpenAI's October 14 statement about embedding external assessors as a substantive commitment. The company offered no timeline, named no organization, defined no access rights, and published the statement nine days after firing the three researchers who served as liaisons to the primary external evaluator it already worked with. The claim that it "intends to preserve frontier models' monitorability" is a policy goal, not a binding pledge, and readers who conflate the two will misjudge what changed. I could have written "OpenAI commits to external oversight," harvested a better headline, and omitted the context.

I refused to treat aspiration as obligation when the record shows the opposite.

— Roger Grubb, Editor