The Financial Times reports that Anthropic withheld its latest AI model from a UK testing agency, limiting external scrutiny of that model’s safety and capabilities.

state: corroboratedheat: lowuncertainty: mediumnovelscott: lowanthropic ai-evaluation model-safetyAnthropic Financial Times

What is this?

Financial Times reporting (reached via echoes and corroborated by FirstPost, IT Pro, AI Weekly and MT Newswires) says Anthropic gave the UK AI Security Institute (AISI) no pre-release access to Claude Mythos 5.1 — the safeguards-stripped variant launched 1 September 2026 for vetted cybersecurity and life-sciences partners alongside the generally available Fable 5.1 — the first AISI exclusion from an Anthropic frontier release; access went to vetted US organisations instead, prompting UK-government fears of a protectionist shift among US labs amid US export-control policy. Related FT headlines in the results ('Anthropic suspends latest AI models after US blocks access to foreigners'; 'UK financial regulators rush to assess risks of Anthropic's latest AI model') show US-government access restrictions and UK regulatory scramble entangled with the episode. A second reporting line in the case (30 Sep) attributes the hold to a White House-directed review covering both OpenAI's and Anthropic's new models, which would reframe vendor refusal as intergovernmental friction. The web results corroborate the withholding event and the US-controls backdrop but do not settle whether the exclusion was Anthropic's unilateral choice or Washington's directive — and the FT primary text remains paywalled.

Why it matters to Scott

The Sep 30 reframing severs the link that grounded the earlier known/low call: a White House-directed hold covering both labs is state-mediated evaluation-access gating, not vendor-controlled verification, so it no longer illustrates Conversion Firewall's proposer-approver split and Scott's canon holds no position on intergovernmental access gating. It is lineage in the radar's Gold Eagle / US-prerelease-review cluster — and usefully dissolves the apparent conflict with the Amodei independent-evaluation commitment — but as another episode of an already-tracked government-gating pattern, with negligible engagement, it would not change what Scott builds or argues.

radar:white-house-gold-eagle-frontier-accessradar:us-open-model-prerelease-testingradar:amodei-frontier-pacing-commitmentradar:concept.model-evaluation

queries asked of Scott's wikis

capability audit independent external model evaluation access
conversion firewall restricted model variants vetted access tiers
frontier release adoption vetting pre-release trust workflow
export controls US government model access open weights
local models sovereignty external scrutiny fallback

Measured heat

now 0 pts/hpeak 6 pts/hcomments 0/hpeers p14momentum: steady2 platformsage 794h

points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion

How the heat travelled

09-08 14:00	⭐ origin echo-reconstructed	The FT reported that Anthropic declined to submit Claude Mythos 5.1 to the UK AI Security Institute before release. Its indexed subtitle say Financial Times on other (echo) · attributed from hn.story.49621024	—
09-09 04:32	first on hacker news · published · +14.6h	Anthropic withheld latest AI model from UK testing agency JumpCrisscross	—
09-09 04:32	amplified on hacker news 👑	hn.story.49621024 JumpCrisscross	peak 5 · 0 comments · 71% of case engagement
09-30 09:15	amplified on hacker news	hn.story.49906340 theanonymousone	peak 2 · 0 comments · 28% of case engagement
09-09 05:21	our radar first saw it · +15.4h	discovery anchor: hn.story.49621024	—

pace: p36 vs 519 stories at the 720h mark (now 794h old) — ahead of addom-local-coding-harness (1.5x), behind checkly-agentic-go-rewrite (0.8x)

Evidence (3) — ⭐ canonical anchor

source	object	author	score	comments
🟧 hn	Anthropic withheld latest AI model from UK testing agency	JumpCrisscross	5	0
🟧 echo.other ⭐	The FT reported that Anthropic declined to submit Claude Mythos 5.1 to the UK AI Security Institute before release. Its indexed subtitle say	Financial Times	—	—
🟧 hn	White House asks OpenAI, Anthropic hold new models from UK testers until review	theanonymousone	2	0

Interpretation history

2026-09-30T09:57:58Z
grounded: novel/low — The Sep 30 reframing severs the link that grounded the earlier known/low call: a White House-directed hold covering both labs is state-mediated evaluation-acces
2026-09-30T09:50:47Z
The second reporting line corroborates that new Anthropic (and OpenAI) models were held from UK testers, but attributes the hold to a White House-directed US review rather than unilateral Anthropic protectionism — reframing the case from vendor-chosen opacity to intergovernmental evaluation-access friction. That shift weakens its fit as a Conversion Firewall illustration and makes the cached grounding's vendor-refusal framing stale, so a reground on the government-review story is warranted.
2026-09-30T09:24:16Z
evidence attached: hn.story.49906340 — Materially transforms the open case: the UK-tester withholding may be a White House-directed US review covering both OpenAI and Anthropic, not a unilateral Anthropic choice.
2026-09-11T06:26:48Z
This review adds no substantive evidence: the alleged exclusion remains credible FT testimony reconstructed through echoes, not independently corroborated reporting. The circumstances and model identity remain unsettled, so the case warrants continued low-temperature monitoring rather than promotion or closure.
2026-09-09T05:27:57Z
The reported first-time exclusion of AISI remains a concrete external-evaluation access concern, but the FT reconstruction and HN link represent one reporting line, not independent corroboration. Nothing new establishes the model identity, circumstances or implications for Scott’s decisions; the previously routed alert needs no repeat.
2026-09-09T05:27:04Z
grounded: known/low — The concern about vendor-controlled access to independent evaluation is already held in Scott’s Conversion Firewall and Capability Audit pages; as supplied, thi
2026-09-09T05:24:22Z
origin walked (codex/luna, conf 0.84): anchor hn.story.49621024 -> echo.other.98775a45a4 by Financial Times
2026-09-09T05:22:38Z
case created — This is a distinct, consequential model-evaluation access episode, but the available evidence is only a reported headline without details or corroboration.

Decision trace

09-30 19:57repriceThe second reporting line corroborates that new Anthropic (and OpenAI) models were held from UK testers, but attributes the hold to a White House-directed US review rather than unilateral Anthropic pr
09-30 19:57groundThe Sep 30 reframing severs the link that grounded the earlier known/low call: a White House-directed hold covering both labs is state-mediated evaluation-access gating, not vendor-controlled verifica
09-30 19:24attachMaterially transforms the open case: the UK-tester withholding may be a White House-directed US review covering both OpenAI and Anthropic, not a unilateral Anthropic choice.
09-30 19:24propose_attachMaterially transforms the open case: the UK-tester withholding may be a White House-directed US review covering both OpenAI and Anthropic, not a unilateral Anthropic choice.
09-11 16:26repriceThis review adds no substantive evidence: the alleged exclusion remains credible FT testimony reconstructed through echoes, not independently corroborated reporting. The circumstances and model identi
09-11 16:26alert_silentThe reported first-time AISI exclusion already received an alert route; the staleness trigger supplies no new consequential delta. Another notification would duplicate attention without changing Scott
09-11 16:26alert_routeThe reported first-time AISI exclusion already received an alert route; the staleness trigger supplies no new consequential delta. Another notification would duplicate attention without changing Scott
09-09 15:27repriceThe reported first-time exclusion of AISI remains a concrete external-evaluation access concern, but the FT reconstruction and HN link represent one reporting line, not independent corroboration. Noth
09-09 15:27alert_silentThe credible reported exclusion already received an alert route, and this review adds neither new evidence nor a material change. Silence avoids duplicate attention without retracting the earlier repo
09-09 15:27alert_routeThe credible reported exclusion already received an alert route, and this review adds neither new evidence nor a material change. Silence avoids duplicate attention without retracting the earlier repo
09-09 15:27alert_shadowThe consequential delta is a reported first-time exclusion of the UK AI Security Institute, not another general concern about vendor-controlled evaluations. Credible FT reporting warrants knowing toda
09-09 15:27alert_routeThe consequential delta is a reported first-time exclusion of the UK AI Security Institute, not another general concern about vendor-controlled evaluations. Credible FT reporting warrants knowing toda
09-09 15:27groundThe concern about vendor-controlled access to independent evaluation is already held in Scott’s Conversion Firewall and Capability Audit pages; as supplied, this is only a possible illustration, not a
09-09 15:24promote_anchororigin walk conf 0.84
09-09 15:22createThis is a distinct, consequential model-evaluation access episode, but the available evidence is only a reported headline without details or corroboration.