UK AISI → Anthropic

1.0.0-rc.1 · Evidence cutoff 12 September 2026

Directly documented interface

Evaluation

Named evaluator performed a model-specific exercise.

evaluator → provider. Evaluator → developer of model actually tested. No inference of joint model development.

Displayed claim → analysis (if present) → coded attributes → interface / instrument → atomic evidence → source version

Instrument and timeline

Instrument: Mythos 5 July cyber-range evaluation
Bindingness: Not applicable to historical event
Announced: Unknown
Observed by: 2026-07-28 · event observed by day; full evaluation period unknown
Effective period: Unknown → Unknown
Event: Conducted
Status at cutoff: historical event only
Last confirmation: 2026-07-28

The incident halted related activity; this qualifies evaluation continuity, not the occurrence of the tests.

Model-specific historical exercise; current access and agreement terms unresolved.

Status investigation

Checked 2026-09-12; assessed as of 2026-09-12. Last confirmation does not establish uninterrupted continuity.

amendments: No change to historical event established
expiry: Historical event, not an access term
termination: Related evaluations halted after incident
restrictions: Non-public configurations and incident-specific response
succession: Same UK AISI unit; no transfer of historical US agreement
conflicts: Reports distinguish incidents from ordinary public deployment

Search queries

site.aisi.gov.uk "OpenAI" "2026" evaluation
UK AISI Anthropic Fable evaluation 2026 suspension status before:2026-09-13

Analytical assessment

No reviewed analytical conclusion is attached. This does not establish independence or absence of an institutional constraint.

Coded attributes

evaluation-capacity: substantial-access-constrained

Ability to design and execute a substantive assessment independently, conditional on access.

Evaluator × evaluation exercise or programme × date. Claude 3.5 Sonnet 2024 pre-deployment exercise; assessed 2024-11-19.

Completed independent tests establish capacity; explicit limited testing access narrows conclusions. Independently run assessments counters inference of provider-directed findings.

Supporting evidence: independent-tests access-limits . Counterevidence: independent-tests

Coding applies only to stated context and date, not a permanent institutional rank.

Published rubric anchors

absent: Affirmative evidence of no ability to perform this assessment.
nascent: Initial methods/team with no completed substantive assessment.
substantial-access-constrained: Completed independent assessment with documented access or time constraints materially limiting conclusions.
substantial-durable-access: Completed assessment and enforceable/reproducible access sufficient for the task over its defined period.

Unresolved when necessary evidence is missing or adjacent anchors cannot be distinguished. No numeric conversion. Absence requires affirmative remit or operational evidence, not missing sources.

technical-control: shared-constrained

Practical ability to alter the model, its availability, or safeguards in the specified deployment. Contract clauses alone are insufficient.

Actor × exact model/deployment × period. Anthropic provision and replacement of Claude models for DoW, contrasted with deployed operational control; 2025–August 2026; assessed 2026-08-27.

Anthropic provides model artifacts and responds to feedback; DoW and cloud providers independently gate operational deployment. Existing deployed models are outside Anthropic technical alteration or shutdown. Consequential functions are divided across institutions, satisfying shared/constrained control without attributing a real-time veto.

Supporting evidence: court-model-provision court-update-gate . Counterevidence: contractual-control court-static-deployment

Coding applies only to stated context and date, not a permanent institutional rank.

Published rubric anchors

little-control: No demonstrated practical control over the specified consequential operation, with affirmative contrary evidence.
shared-constrained: Consequential controls divided among actor, host, customer or enforceable technical constraints.
substantial: Actor demonstrably retains major development, update or access controls in this environment.
dominant-lifecycle: Actor demonstrably controls development, deployment, updates and access across the specified lifecycle without a consequential independent controller.

Unresolved when necessary evidence is missing or adjacent anchors cannot be distinguished. No numeric conversion. Absence requires affirmative remit or operational evidence, not missing sources.

public-authority: limited

Institutionally authorized power to make decisions binding in the specified context.

Institution × legal/operational context × assessed date. February 2024 stated evaluation and advisory remit; assessed 2024-02-09.

Self-described supplementary evaluation and policy-advice remit supports advisory anchor. It does not code the powers of DSIT or UK government as a whole.

Supporting evidence: aisi-remit . Counterevidence:

Historical institutional remit. Later capacity, legislative changes or government powers require separate evidence. Limited denotes advisory remit here, not limited binding authority.

Published rubric anchors

absent: Affirmative remit excludes binding public decisions in this context.
limited: Advisory or delegated bounded public decisions.
substantial: Binding decisions over a significant but bounded function.
binding-operational: Documented binding decision and ability to carry it into operation in the defined function.

Unresolved when necessary evidence is missing or adjacent anchors cannot be distinguished. No numeric conversion. Absence requires affirmative remit or operational evidence, not missing sources.

information-access: unresolved

Ability to obtain non-public model information independently of provider consent.

Institution × information category × jurisdiction/date. Compulsion of access to non-public frontier checkpoints; assessed 2026-09-12.

A limited voluntary exercise cannot prove absence of all compulsory authority. No jurisdiction-wide classification from this evidence.

Supporting evidence: access-limits . Counterevidence: independent-tests

Coding applies only to stated context and date, not a permanent institutional rank.

Published rubric anchors

absent: Affirmative evidence excludes access in this context.
voluntary-provider-controlled: Access obtained through provider permission or negotiated instrument.
partial-compulsory: Specific compulsory information powers apply to some requested information.
broad-compulsory: Applicable legal powers reach the information needed for this defined task, subject to stated safeguards.

Unresolved when necessary evidence is missing or adjacent anchors cannot be distinguished. No numeric conversion. Absence requires affirmative remit or operational evidence, not missing sources.

Atomic evidence and exact sources

The evaluation had a limited access and testing period.

Source: Claude 3.5 Sonnet joint testing report · US AISI / UK AISI
Version: 2024-11-19; retrieved 2026-09-12
Exact locator: Section 2.1, printed pages 1–2
Limit: Applies to this exercise; does not identify who caused the limits.

Retrieved bytes SHA-256: cee68f384975897e9ae8fd556baa9ebae9919fe08fe1e1abb363cbead7802903

The institutes each ran independent tests and worked together on methodology and interpretation of findings.

Source: Claude 3.5 Sonnet joint testing report · US AISI / UK AISI
Version: 2024-11-19; retrieved 2026-09-12
Exact locator: Section 1, printed page 1
Limit: Independence of tests does not establish durable model access.

Retrieved bytes SHA-256: cee68f384975897e9ae8fd556baa9ebae9919fe08fe1e1abb363cbead7802903

Anthropic could not technologically enforce contractual usage restrictions and lacked direct visibility into DoW use.

Source: Anthropic v DoW: summary judgment, document 250 · US District Court, Northern District of California
Version: 2026-08-27; retrieved 2026-09-12
Exact locator: Printed page 6
Limit: Specific deployed environment; does not imply no control over future development, contracts or provision.

Byte capture unavailable. HTTP Error 403: Forbidden Published document identity and locator remain recorded.

In its February 2024 approach document, AISI describes itself as a secondary check rather than a regulator.

Source: AI Safety Institute approach to evaluations · DSIT / AI Safety Institute
Version: published-2024-02-09; retrieved 2026-09-12
Exact locator: A note on evaluations, final paragraph
Limit: Historical stated remit; not an assessment of every governmental power or all powers at September 2026.

Retrieved bytes SHA-256: 3cf1fee953c6819e009c82b58e5340b7e67737a824477c3ceced9ca636be981c

The court describes deployed DoW Claude models as static and beyond Anthropic technological access, alteration or shutdown.

Source: Anthropic v DoW: summary judgment, document 250 · US District Court, Northern District of California
Version: 2026-08-27; retrieved 2026-09-12
Exact locator: Printed page 17, Section II.F, first paragraph
Limit: Applies to deployed models described in this litigation, not future model provision or hosted consumer services.

Byte capture unavailable. HTTP Error 403: Forbidden Published document identity and locator remain recorded.

The court records that replacement Claude models undergo third-party cloud-provider and DoW security testing and evaluation before operational approval.

Source: Anthropic v DoW: summary judgment, document 250 · US District Court, Northern District of California
Version: 2026-08-27; retrieved 2026-09-12
Exact locator: Printed page 17, Section II.F, first paragraph
Limit: Identifies a deployment approval gate, not complete interpretability, universal evaluation sufficiency or provider substitutability.

Byte capture unavailable. HTTP Error 403: Forbidden Published document identity and locator remain recorded.

The court records Anthropic providing models to DoW or a primary defense contractor and responding to DoW evaluation feedback and requests.

Source: Anthropic v DoW: summary judgment, document 250 · US District Court, Northern District of California
Version: 2026-08-27; retrieved 2026-09-12
Exact locator: Printed page 17, Section II.F, first paragraph
Limit: Provision is distinct from control over deployed inference; future supply terms and update alternatives are unspecified.

Byte capture unavailable. HTTP Error 403: Forbidden Published document identity and locator remain recorded.

AISI identifies Mythos 5 as a model it tested in its July 2026 cyber-range exercise.

Source: Incident Report: unsanctioned agent behaviour during cyber testing · UK AI Security Institute
Version: consulted-2026-09-12; retrieved 2026-09-12
Exact locator: Paragraphs beginning The incident stemmed and Almost all of this behaviour
Limit: Specific non-public testing configurations; not ordinary deployed behavior or proven real-world harm.

Retrieved bytes SHA-256: d877ecd7c74dcf9b667aadb79c49e784e11f4a87eda89f287089038b8584350c

OpenAI reports AISI stopped related evaluations and isolated relevant machines after July28 incident detection.

Source: Third-party cyber evaluations involving OpenAI models · OpenAI
Version: consulted-2026-09-12; retrieved 2026-09-12
Exact locator: UK AISI section, paragraph beginning UK AISI identified the activity
Limit: Specific exercise response; not termination of all developer–institute relations.

Byte capture unavailable. HTTP Error 403: Forbidden Published document identity and locator remain recorded.

Supporting and conflicting evidence

Interface support: uk-anthropic-july-tested. Counterevidence: july-evaluation-halted

Substantive review

evidence/access-limits: approved · agent reviewer codex-research-checker · 2026-09-12

Full joint report section 2.1 explicitly limits access and the testing window. It does not identify the institution responsible for either constraint. Checked locator: Section 2.1, printed pages 1–2. Source version: 2024-11-19.