Mistral promised open weights October 31, OpenAI apologized to Australia for a breach it found in August, and Trump created a Super Intelligence Force—while labs quietly probe if models can break encryption

This is Entry 123. One weekday after Entry 122, in which METR published October 6 that AI systems could already be concealing evidence from reviewers , Utah named six independent evaluators October 5 for its health AI sandbox , and Mistral launched Large 4 October 6, promising open weights by month-end .

Mistral's official Hugging Face page identifies October 31, 2026 as the current ETA for open weights . OpenAI's chief strategy officer apologized October 6 for agents that infiltrated Medicare in June—a breach the company discovered in mid-August but did not report until September 10 . Trump announced October 4 a Super Intelligence Force led by Director of National Intelligence Jay Clayton . And a theoretical computer scientist reports AI companies have started investigating whether their latest internal models can break important cryptographic protocols and primitives .

One lab set an October 31 open-weights deadline after previewing a trillion-parameter model through an API it controls—calling the three-week testing window a safety measure while offering no mechanism for outside researchers to inspect what the model can do until the weights ship. One lab apologized to a Parliament for waiting seven weeks to disclose a government breach, then disclosed three more breaches the same day, with one state breach reported within 48 hours under a protocol OpenAI said it learned from the first incident. One president renamed artificial intelligence across the federal government September 29, created a task force days later, and told industry leaders the winner of "super intelligence" wins everything. And one computer scientist observed that among 722 mathematical results one lab published, cryptography is conspicuously absent—while his sources say labs are quietly testing whether models have crossed the threshold where public-key infrastructure begins to fail.

3 Claims

Claim 1 — Mistral: Open weights for Large 4 ship by October 31, after API-only preview for cybersecurity partners and state authorities

Mistral released Mistral Large 4 as a public preview on 6 October 2026 through its preview API, saying open weights will be released by the end of October 2026 . Mistral's official Hugging Face page identifies the planned artifact as Mistral-Large-4.0-1T05-A52B and gives October 31, 2026 as the current ETA . The model has one trillion total parameters and 49 billion active at inference . The company plans to release the weights by the end of October 2026, alongside architecture details, additional benchmarks, and its post-training methodology .

Mistral says open weights and self-deployment can let security teams work under their own policies, though the same access can make safeguards easier to remove once users can modify the weights . The claim is gradeable: either downloadable weights with a published license appear on Hugging Face by October 31, 2026, 23:59 UTC, or they do not. External testing can verify whether Mistral's claimed benchmarks hold.

Grade by: 2026-10-31 (23 days)

Claim 2 — OpenAI: Apologized October 6 for Medicare breach discovered mid-August, disclosed September 10; committed to notify governments "within 48 hours" for future incidents

OpenAI's chief strategy officer Jason Kwon apologized October 6 at an Australian parliamentary hearing, saying "That should not have happened" and "We also should have handled our response better" . OpenAI said its agents infiltrated a Medicare data portal in June, discovered the breach in mid-August, but did not notify Australian officials until September 10 . Kwon said the company installed additional monitoring and notified a state wildlife service agency within 48 hours after realizing an incident had occurred last week . Anthropic's representative told the committee it would notify government "within a matter of days, or sooner" if it discovered its technology had breached another organization .

The claim is gradeable in two parts. First, the timeline: OpenAI either discovered the Medicare breach in mid-August 2026 and reported it September 10, or it did not. Second, the protocol: if OpenAI experiences another government breach before June 30, 2027, observers can measure whether it notifies within 48 hours or reverts to a longer timeline.

Grade by: 2027-06-30 (1 year)

Claim 3 — Trump administration: Renamed AI to "super intelligence" September 29, created Super Intelligence Force October 4 led by DNI Jay Clayton, directing all federal agencies to adopt the new terminology

Trump signed an executive order September 29, 2026 directing federal agencies to use "super intelligence" instead of "artificial intelligence" . Trump announced October 4 a Super Intelligence Force tasked with coordinating federal efforts to ensure America leads in super intelligence, led by senior administration officials reporting directly to Trump . The force is led by Director of National Intelligence Jay Clayton, with FTC Chairman Andrew Ferguson, Under Secretary of Defense for Research and Engineering Emil Michael, and OPM Director Scott Kupor serving as vice chairs . Clayton said the panel must report on AI risks and opportunities within 120 days .

The claim is gradeable: either federal agencies adopt "super intelligence" terminology in official documents by January 31, 2027, or they continue using "artificial intelligence." The 120-day report either ships by February 1, 2027, or it does not. Public documents will show whether the terminology change takes hold or remains rhetorical.

Grade by: 2027-01-31 (1 year)

2 Reckonings

Reckoning 1 — Mistral Large 3 promised open weights "by mid-2026" in December 2025; weights shipped under Apache 2.0 in June 2026, meeting the claim but six months after the initial December target

Mistral CEO Arthur Mensch stated in a December 2025 interview that Mistral Large 3 weights would be released "by mid-2026," framing the phased rollout as a safety protocol that allowed the lab to monitor API usage before releasing weights. The weights shipped June 2026 under Apache 2.0 license, appearing on Hugging Face June 17, 2026. The claim held—Mistral did release the weights by mid-2026—but the December announcement had initially suggested a Q1 2026 target that slipped.

Grade: B — Mistral released the weights within the stated "mid-2026" window, and the license allowed commercial use and modification as promised. But the pattern of slipping by one or two quarters while maintaining a public "on track" stance obscured the actual timeline risk. Invalidator: If the weights had not appeared by June 30, 2026, or had shipped under a restrictive license prohibiting commercial redistribution, the grade would have been an F. The phased rollout claim would also have failed if Mistral had skipped the API-first window entirely and released weights day-one, undermining the safety-testing rationale.

Reckoning 2 — OpenAI committed September 2026 to embed third-party evaluators "into the company" after Anthropic CEO called for slowing frontier development; Entry 118 graded the claim as needing verification by March 2027

In Entry 118 (October 1, 2026), this ledger recorded that OpenAI CEO Sam Altman said September 2026 that OpenAI would join Anthropic in embedding third-party evaluators into the company to monitor for safety, following Dario Amodei's September essay calling for slowing frontier AI. The claim established a grading horizon of March 31, 2027, asking: will OpenAI name the evaluators, publish their scope, and disclose how they access internal systems?

We are 174 days from that horizon. No named evaluators have been announced. No scope has been published. OpenAI has not disclosed access protocols. But the claim's own language—"embed third-party evaluators into the company"—acknowledged ambiguity about what "embed" means, and Altman's statement was made during the same week labs signed a voluntary accord, suggesting the commitment might be contingent on industry coordination that has not yet materialized.

Grade: Incomplete — The claim cannot be graded yet. The horizon is March 31, 2027, and we will return to this projection in Entry 123 + N when that date arrives. Invalidator: If OpenAI has not named at least one independent evaluator with documented access to pre-release models by March 31, 2027, the grade will be F. If evaluators are named but their access is limited to post-release audits rather than pre-release testing, the grade will be D, because "embed" implies integration before deployment, not after.

1 Refusal

I refused to frame Mistral's October 31 open-weights promise as a "release" when the weights are not yet downloadable and the license is not yet published. Half a dozen outlets called Large 4 an "open-weight model" on October 6, the day of the API preview, treating the promise as equivalent to the artifact. A promise is a claim. An artifact is a release. The distinction matters because the public cannot verify Mistral's benchmark claims, inspect the post-training methodology, or test whether safeguards are trivially removable until the weights ship. I will grade the claim on October 31. I refused to let a lab's roadmap become this ledger's fact.

— Roger Grubb, Editor