Microsoft claims 83% diagnostic accuracy, UK holds senior managers liable for AI, and Railway says attackers may have decrypted secrets—8 days before Mistral's weights deadline

This is Entry 133. One weekday after Entry 132, in which OpenAI credited GPT-5.6 Sol with finding critical PHP vulnerabilities affecting 70% of websites, Google DeepMind research engineer Sofia Liguori resigned over military AI contracts, and Anthropic opened Claude Duet, a proactive agent that messages users first.

Microsoft published a peer-reviewed study in Science on October 22 claiming its MAI-DxO diagnostic system reached 83% accuracy on complex medical cases, compared with 36% for generalist physicians using reference materials . The UK Financial Conduct Authority published guidance on October 19 stating it will not introduce new AI rules but expects senior managers to be personally accountable for AI systems, warning that too many firms are under-investing in governance . Railway disclosed October 21 that it detected anomalous activity and an unauthorized third party may have accessed encrypted customer environment variables and obtained decrypted contents for some of them . And Mistral Large 4, announced in public preview October 6 with a self-imposed end-of-month release window, remains unpublished as of October 23—Day 22 of 31 .

One lab published a diagnostic benchmark on historical teaching cases and called it ready for real-world evaluation, without naming a date. One regulator told 86% of financial firms they already have the rules they need and named the accountable person. One hosting platform told every customer to rotate every secret after attackers may have decrypted what was stored encrypted at rest. Mistral's October 31 open-weights deadline holds with 8 days remaining.

3 Claims

Claim 1 — Microsoft: MAI-DxO reached 83% accuracy on NEJM cases; ready for real-world clinical evaluation

Microsoft AI published MAI-DxO in Science on October 22, 2026, reporting 83.2% accuracy on 304 NEJM teaching cases when running on GPT-5, against 36.1% for physicians who could use reference tools . MAI-DxO coordinates a virtual physician panel that gathers evidence before committing to a diagnosis . Dominic King, vice president of Health at Microsoft AI, said the technology is ready for real-world evaluation and called it potentially the biggest study of its kind , though no commercial release date has been announced .

Microsoft's own chart reports GPT-5 alone scored 78.9% on the complex cases, with the paper attributing roughly four additional percentage points to the orchestration approach . The study reports physician accuracy of 22% without reference resources and 36% with them on the complex cases , but the trial notes such complex cases are typically managed in clinical practice by specialist physicians or multidisciplinary teams, and specialists were not included .

Claimant: Microsoft
Claim: MAI-DxO ready for real-world clinical evaluation; diagnostic accuracy on live patient cases will match or exceed the 83% reported on historical NEJM teaching cases
Grade by: 2027-04-23 (6 months) — verify whether Microsoft or a clinical partner publishes peer-reviewed results from prospective patient trials showing ≥80% diagnostic accuracy and whether those trials involve specialists or multidisciplinary teams as comparators

Invalidator: If Microsoft does not publish prospective clinical trial results by April 2027, or if published trials show accuracy below 70% when evaluated against specialist teams rather than solo generalists, the claim receives a C.

Claim 2 — UK FCA: Senior managers personally accountable for AI under existing rules; 86% of firms now use AI

The UK Financial Conduct Authority said October 19 it will not introduce new AI rules but expects senior managers to be personally accountable for AI systems, with Jessica Rusu citing a joint FCA and Bank of England survey showing 86% of firms now use AI . The FCA states senior managers are expected to understand AI risks within their areas of responsibility rather than act as technical experts, but warns oversight remains an area where too many firms are under-investing .

The regulator says firms deploying AI safely generally share proportionate risk-based governance, clearly designed human oversight, and senior managers who are accountable for AI systems and for taking reasonable steps to mitigate risks . The publication creates no new requirements and prescribes no single governance model but provides a practical reference point for firms assessing their AI arrangements .

Claimant: UK Financial Conduct Authority
Claim: Existing senior manager accountability framework (SM&CR) sufficient to hold individuals liable for AI harms; no AI-specific regulation needed through at least 2027
Grade by: 2027-10-19 (1 year) — verify whether FCA brought enforcement action against a named senior manager for AI-related harm under SM&CR, or whether UK introduced standalone AI regulation for financial services

Invalidator: If the UK government introduces a financial-services AI bill by October 2027, or if the FCA states publicly that SM&CR proved insufficient for AI accountability, the claim receives a C.

Claim 3 — Railway: Attacker may have decrypted customer environment variables; impact may be limited to subset of users

Railway disclosed October 21, 2026 that it detected anomalous activity and an unauthorized third party may have accessed encrypted environment variables and obtained decrypted contents for some of them . Railway's initial analysis suggested the impact may be limited to a subset of users, but it told every customer to replace all secrets stored on its platform . An on-chain investigator reported crypto keys stored on Railway were drained, about $371,000 on October 21, across 500+ transactions and 200+ addresses .

Railway paused all deployments, revoked and reissued storage credentials, signed some users out, and said it would email the customers it confirms were affected . Users said the notice came roughly 15 hours after the first signs of trouble . Railway's terms cap liability at $100 or one year's fees .

Claimant: Railway
Claim: Impact of October 21 breach limited to subset of users; attacker access to decrypted environment variables did not reach majority of Railway's customer base
Grade by: 2026-11-21 (1 month) — verify whether Railway publicly states what percentage of users were affected, and whether that percentage exceeds 50%

Invalidator: If Railway discloses or credible third-party analysis establishes that more than 50% of active users had environment variables accessed or decrypted, the claim receives a C.

2 Reckonings

Reckoning 1 — Mistral's October 31 open-weights deadline for Large 4: 8 days out, weights still unpublished, license terms unknown

Original claim: Mistral AI announced October 6, 2026 that open weights for Mistral Large 4 would be released by the end of October, with downloadable weights to follow by the end of the month according to the Paris-based company's launch post .

What happened: As of October 23—Day 22 of 31—the weights and commercial license remain unpublished, with Hugging Face model card fields for both weights and license still reading "Coming soon" . Industry reports suggest Mistral is planning to move away from the permissive Apache 2.0 license used for Mistral Large 3, opting instead for a custom framework that would align Mistral with a $20 million monthly revenue gate for commercial use .

Grade: B — Mistral delivered a public preview API on the announced date but has not delivered the promised open weights with 8 days remaining in the stated window. The license terms that will govern "open" remain undefined. A final grade awaits October 31.

Invalidator: If weights and license ship by October 31 with Apache 2.0 or equivalent permissive terms, the claim receives an A. If weights do not ship by October 31, or ship under a license with revenue-based commercial restrictions not disclosed in the October 6 announcement, the claim receives a C.

Reckoning 2 — OpenAI's October 1 firing of three safety researchers who liaised with METR: promised embedded external assessors 13 days later

Original claim (from Entry 128): OpenAI announced October 1 that it fired three AI safety researchers for violating rules on handling sensitive information, with Tomek Korbak's role involving speaking with METR and serving as OpenAI's technical point of contact for METR and Redwood Research during the Hugging Face investigation . OpenAI said October 14 it will embed external assessors and work closely with independent safety observers .

What happened: OpenAI disclosed October 21 that AI-discovered PHP flaws triggered emergency patches and credited GPT-5.6 Sol with identifying a critical flaw , and OpenAI made its Decisions API generally available October 22 , but the company has not published details of embedded external assessors, their institutional affiliations, or their access permissions as of October 23—nine days after the October 14 commitment.

Grade: C — OpenAI fired the researchers who served as institutional bridges to METR and Redwood, promised embedded external assessors 13 days later, then disclosed a major vulnerability discovery and launched a new API product without publishing the structure, names, or access protocols for those assessors. A commitment to embed observers means nothing if the observers remain unnamed and their mandate undefined two weeks later.

Invalidator: None — the October 14 promise was to "embed external assessors" and work "closely with independent safety observers." By October 23, no assessor names, no institutional home, no defined access, no monitoring protocol. The promise remains a press release.

1 Refusal

This morning I had a draft that opened with "Black October cybersecurity fears grip internet infrastructure" and cited coverage calling October 2026 one of the roughest stretches for internet security in recent memory. I deleted the draft. The phrase "Black October" appeared in sensationalist aggregator coverage, not in primary sources from Railway, LiteLLM, OpenCode, PHP, or any national CERT. The security incidents this week are real—Railway told every user to rotate every secret, OpenCode disclosed a distillation attack, and PHP patched two critical remote-code-execution flaws—but I will not borrow a catastrophe label invented by outlets chasing traffic. This ledger names the operator, states the claim, cites the source, and sets the grading horizon. It does not package incidents under a brand.

I refused to use a third-party label that repackaged discrete incidents as a named event when no operator or regulator used that term.

— Roger Grubb, Editor