DELOITTE (AI ASSURANCE AND TESTING)

OLYMPUS opened an institutional file on Deloitte’s AI assurance and testing work, not on Deloitte. The firm is a 470,000-person professional-services network and this file does not try to profile it. It covers the narrow part that appears in the frontier labs’ own documents: the team that red-teams models for biological risk and designs the uplift trials, the practice that sells AI controls and assurance, and the federal and UK contracts that carry the word AI in their descriptions. This is not a psychometric profile, since an institution has no Dark Triad. It is a scope, a money record, a customer list and a voice. The finding is the arrangement: the firm whose graded trial and red team sit under Anthropic’s ASL-3 decision for Claude Opus 4 is also the firm that put Anthropic’s model in 470,000 hands.

Evaluator record: the engagements, the terms each document states, the findings quoted and what followed.

Institutional Archetype

THE AUDITOR WITH AN INTEGRATOR’S BADGE — Deloitte tests frontier models for three developers’ release documents and sells assurance services to enterprises, and in October 2025 it joined Anthropic in announcing “Anthropic’s largest enterprise AI deployment to date”. Audit firms have always sold the audit and the adjacent consulting. What is new is that the object under test is a product of a company whose model the same firm deploys to its own people and builds client solutions on, and the tester’s documents are the ones the lab’s safety level cites. The public record shows no impropriety. It also shows no independence statement for this practice.

Mandate & Origin

Scope. Deloitte is a network. Anthropic’s announcement carries the standard footnote: “Deloitte refers to one or more of Deloitte Touche Tohmatsu Limited (DTTL), its global network of member firms, and their related entities (collectively, the ‘Deloitte organization’). DTTL (also referred to as ‘Deloitte Global’) and each of its member firms and related entities are legally separate and independent entities” (Anthropic, 6 October 2025). The testing named in the labs’ documents is by Deloitte Consulting LLP in the United States; the assurance pages read here are Deloitte US and Deloitte UK; the defence award is to Deloitte LLP.

The biosecurity team’s origin. Deloitte’s own page: “In late April 2024, Deloitte acquired substantially all the assets of Gryphon Scientific, LLC (Gryphon), a leader in biosafety, biosecurity, and all-hazards preparedness and response, with market-leading experience in using AI to enhance security and safety.” It adds “Gryphon pioneered AI safety, charting how AI can change the risk landscape of biological and chemical threats,” and that Gryphon’s experts “will advance Deloitte’s capabilities in life science and public health preparedness and response, catalyzing the growth of Deloitte’s research, biosecurity, laboratory safety, and emergency preparedness and response capabilities” (Deloitte, Deloitte Acquires Gryphon Scientific). Amazon’s Nova report names the continuity: “The evaluation team at Deloitte Consulting LLP (formerly known as Gryphon Scientific) has unique experience at the intersection of artificial intelligence and biology” (Amazon, Nova technical report). Rocco Casagrande, Gryphon’s founder, is listed as “Managing Director for Deloitte Consulting” (Georgetown Journal of National Security Law and Policy). The firm’s earlier history as an evaluator is on the Gryphon Scientific profile.

The assurance practice, in its own words. Deloitte US, 12 August 2026: “Deloitte today announced expanded AI Controls and Assurance services and solutions designed to help organizations confidently adopt, scale and govern AI across the enterprise,” covering “AI risk assessments, model validations, and redesigning controls and processes — particularly where AI impacts financial reporting and operational integrity.” It says the capabilities are “developed in alignment with its Trustworthy AI™ framework” (Deloitte US, press release). Deloitte UK’s AI Assurance page: “Providing objective third-party challenge over the technical build and performance of your AI use cases,” across organisational, legal and regulatory, model, and platform pillars, citing “global standards such as ISO 42001” and “the EU AI Act” (Deloitte UK, AI Assurance). Its Trustworthy AI page says assurance “often involves independent model testing and/or an assessment of how an AI risk management framework has been implemented and is operating” (Deloitte UK, Trustworthy AI). On 24 September 2026 Deloitte reported FY2026 global revenue of “US$74.5 billion” and “a US$3 billion investment through 2030 to modernize delivery and operations across all our businesses” (Deloitte, FY2026 revenue).

Funding & Backers

Revenue and customers. Deloitte states FY2026 revenue of US$74.5 billion, “a 3.8% increase in local currency and 5.7% increase in USD from FY2025” (Deloitte, FY2026 revenue). No lab document and no Deloitte page states a fee for any frontier-model testing, a contract value or whether any of it was paid. What the pages state is commercial alignment with the labs it tests:

• Anthropic. On 6 October 2025: “Anthropic and Deloitte today announced an expanded alliance that will make Claude available to Deloitte people across its global network and develop new industry-specific solutions powered by Claude,” “The collaboration represents Anthropic’s largest enterprise AI deployment to date, available to more than 470,000 Deloitte people,” with a Claude Center of Excellence and certification of 15,000 professionals. Deloitte Global’s Ranjit Bawa: “Deloitte is making this significant investment in Anthropic’s AI platform because our approach to responsible AI is very aligned” and “Anthropic shares our passion for safety and reliability” (Anthropic, 6 October 2025). Deloitte’s alliance page: “Deloitte’s holistic approach, anchored in its Trustworthy AI™ (TAI) Framework, integrates seamlessly with Anthropic’s Constitutional AI (CAI) deployment” (Deloitte US, Anthropic alliance). See Anthropic.
• Meta. Mark Zuckerberg’s 23 July 2024 letter: “Companies like Scale.AI, Dell, Deloitte, and others are ready to help enterprises adopt Llama and train custom models with their own data” (Meta, Open Source AI Is the Path Forward). See Meta and Mark Zuckerberg.

The government as customer, US. A USAspending award search over contracts with a recipient named Deloitte and “artificial intelligence” or related terms in the award description returns these awards (obligations as USAspending prints them; FY2026 awards are in progress):

Recipient names are “DELOITTE CONSULTING LLP” except where marked; USAspending prints the parent as “DELOITTE FINANCIAL ADVISORY SERVICES LLP” on most and “DELOITTE TOUCHE TOHMATSU LIMITED” on the Commerce award. The CMS order’s ceiling with all options is $28,234,945.08 and the NIH order’s $8,484,758.71 (USAspending: NIH; CMS; Fiscal Service; Commerce; PBGC; FinCEN GenAI; NTSB, 2025; NTSB, 2026; Navy; Labor). Gryphon-era federal work that Deloitte’s page lists: for the Defense Threat Reduction Agency “the Gryphon team developed the Disease Outbreak Risk Assessment (DORA) tool,” and for CDC’s NIOSH it “screened nearly 200,000 publication titles” (Deloitte, Gryphon page).

The government as customer, UK. The Ministry of Defence’s contract award notice “AI Technical & Data Science Support” (published 16 December 2025, notice 2025/S 000-083515): “supply of client-side AI Technical support for Defence AI Centre (DAIC) Capability Development programmes,” five lots, with “Lot 2 - AI Assurance” listed against “Deloitte LLP,” and Capgemini UK Plc, Applied Data Science Partners, Rowden Technologies and Advai on the other lots; “Number of tenders received: 49”; the contract concluded 16 December 2025 with “Total value of the contract/lot: £17,500,000” for the whole (Find a Tender). The notice states no per-lot value.

Institutional Voice & Intent

Deloitte US, 12 August 2026, Frank Milano, U.S. Assurance leader: “Organizations are deploying AI systems that make critical decisions, but the gap between ambition and assurance is widening. Deloitte helps organizations close the gap by assessing what their AI actually does and developing governance that enables innovation instead of blocking it.” Sarah Fedele, Assurance Strategy and Transformation Leader: “Controls and assurance is the infrastructure that makes AI trustworthy at scale.” The release cites Deloitte’s own survey: “While 74% of companies plan to deploy agentic AI within two years, only 21% report having a mature governance model for autonomous agents” (Deloitte US, press release).

Stated intent: Help organizations “adopt, scale and govern AI” with assurance, controls and, on the frontier-model side, the biosecurity and AI-safety expertise it bought in 2024.

Observed intent: Be the tester and the integrator at once. It appears in release documents as a tester of Anthropic, Meta and Amazon models, sells assurance of enterprise AI, deploys Claude to its own staff and builds client solutions on it, and holds the UK MoD’s AI Assurance lot.

Gap: The assurance pages describe “objective third-party challenge” and the Anthropic announcement quotes Deloitte Global on “this significant investment in Anthropic’s AI platform” and an alliance. Both are Deloitte’s statements about different business lines of the same network. The record does not settle whether the testing team and the alliance team are separated or what the labs pay for testing; neither Deloitte nor Anthropic states a firewall or a fee. For the arrangement it does not need to (Deloitte UK, AI Assurance; Anthropic, 6 October 2025).

Interest: Deloitte argues for practices in its own words; what adoption would pay it is set out from its own pages and the documents above. No fee from any lab is stated.

• Third-party AI assurance as the trust mechanism. Position: “Third-party AI assurance has a vital role to play in building trust in AI systems; this has been recognised by the UK government as a critical enabler for the UK to realise the full potential of AI.” What it pays: assurance is the product: “expanded AI Controls and Assurance services and solutions,” “AI risk assessments, model validations,” and the MoD’s AI Assurance lot; a rule that requires independent assurance of AI systems creates the demand its practice sells into. Best case, Deloitte’s: “the gap between ambition and assurance is widening,” with 21 percent of companies reporting a mature governance model for autonomous agents. Convergence: the UK government, which Deloitte cites; the standards bodies whose ISO/IEC 42001 and EU AI Act conformity requirements it names as its frame; and the audit-and-certification side of the evaluator layer, where AIUC’s consortium list prints Deloitte staff (“Partner & EMEA CISO at Deloitte,” “CyberRisk Leader at Deloitte,” “Associate Director at Deloitte”) (Deloitte UK, AI Assurance; Deloitte US, press release; AIUC-1, consortium). See AIUC.
• Standing biosecurity evaluation of AI models. Position: Gryphon “helped shape AI safety,” and Deloitte and Gryphon “will work together to develop practical AI applications in health, foster collaboration, and identify and manage risks to produce reliable AI solutions.” What it pays: the evaluation team is a growth line (“catalyzing the growth of Deloitte’s research, biosecurity, laboratory safety”), and each lab document that names Deloitte is a delivered service; Anthropic’s cards reuse its trial design across models. Best case, Deloitte’s: Gryphon “pioneered AI safety, charting how AI can change the risk landscape of biological and chemical threats.” Convergence: the labs, whose responsible-scaling decisions cite the work, and the other biology evaluators named beside it, SecureBio and Signature Science (Deloitte, Gryphon page).

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

• Social accounts and feeds. No Deloitte account is specific to this practice; Deloitte’s corporate newsroom is deloitte.com/us/en/about/press-room.

Position in the Apparatus

The documents that name it. Each row is the developer’s own text; “Deloitte” and “Deloitte Consulting” are printed as the document prints them.

The decision that cites it. Anthropic’s Opus 4 card: “Participants who had access to Claude Opus 4 scored 63% ± 13%,” the uplift “was 2.53× and 1.70×,” and “When we ran this trial for Claude Sonnet 3.7 during our previous round of testing, Deloitte’s updated rubric placed Sonnet 3.7’s uplift at 1.53×,” concluding that “Claude Sonnet 4 does not meet the ASL-3 bar, while Claude Opus 4’s result is sufficiently close that we are unable to rule out ASL-3.” On the expert red team: “For Claude Opus 4, red-teamers noted substantially increased risk in certain parts of the bioweapons acquisition pathway” (Anthropic, Claude Opus 4 and Sonnet 4 system card). The grading rubric and the red team are Deloitte’s, by the card’s account, and the safety-level conclusion cites both.

The newest cards. Anthropic’s Claude Opus 5.5 card (22 September 2026) says two evaluations “replace two long-form virology task evaluations we have reported in previous system cards. These previous evaluations became difficult to interpret due to saturation of task performance by recent models.” Its Sonnet 5.5 (28 September 2026) and Haiku 5.5 (7 October 2026) cards do not name Deloitte; the Haiku card says “We did not conduct expert red-teaming sessions, uplift trials, or other resource-intensive evaluations requiring human participants” (Anthropic, Opus 5.5; Anthropic, Haiku 5.5).

Under the old name. OpenAI’s GPT-5.5, GPT-5.6 and GPT-6 Astra pages still describe a multiple-choice dataset built with “Gryphon Scientific” that “has not been published,” and name Deloitte nowhere; the firm that built it is the one Amazon calls “Deloitte Consulting LLP (formerly known as Gryphon Scientific)” (OpenAI, GPT-6 Astra; OpenAI, GPT-5.6; Amazon, Nova report).

The adjacency to name without a verdict: one network is the biology tester in a lab’s safety-level decision, the lab’s largest enterprise deployment, a UK defence AI-assurance supplier and a vendor of AI-assurance services. Not a cabal. A firm with a tester’s name tag on one side of the table and an alliance badge on the other, with nothing printed between.

Actions & Leadership Choices

• April 2024. Deloitte acquires substantially all the assets of Gryphon Scientific.
• February 2025 to September 2026. Named in fifteen Anthropic system cards.
• 6 October 2025. Anthropic and Deloitte announce the expanded alliance and the rollout of Claude to Deloitte staff (Anthropic).
• 16 December 2025. UK MoD concludes the AI Technical & Data Science Support award; Deloitte LLP holds Lot 2, AI Assurance (Find a Tender).
• 12 August 2026. Deloitte US announces expanded AI Controls and Assurance services (Deloitte US).
• 24 September 2026. FY2026 revenue and the US$3 billion investment through 2030 (Deloitte).
• September to October 2026. Anthropic’s Opus 5.5, Sonnet 5.5 and Haiku 5.5 cards stop naming it.

CONDUCT: STRUCTURAL — THE FIRM THAT GRADES THE MODEL’S BIOLOGICAL RISK IS ALSO THE MODEL MAKER’S LARGEST ENTERPRISE DEPLOYMENT AND ALLIANCE PARTNER. Both halves are in the parties’ words: the Opus 4 card credits Deloitte’s rubric and red team in its ASL-3 decision, and the 6 October 2025 announcement calls the alliance “Anthropic’s largest enterprise AI deployment to date.” No page says the testing was affected. The point is that the arrangement has no independence statement printed with it, and the labs’ documents do not say what the testers were paid.

Alliance Map

These run at the same time. One network tests Anthropic’s models for biological risk, deploys Claude to its staff and builds client solutions on it, designs the trial a safety level cites, holds a UK defence AI-assurance lot, sells assurance services and bills US agencies for AI projects. Each tie is stated by Deloitte, a lab, a procurement notice or USAspending. The overlap is the finding.

Reach Assessment

Institutional: Its trial design and red team sit inside a frontier lab’s safety-level decision, and its assurance practice and defence lot sit inside government and enterprise AI adoption. Memetic: Almost none: no public announcement of its testing work exists on its own pages; the labs’ cards carry it. Civilizational: A biological uplift trial run and graded by one firm decides whether a model is released under the strictest safeguards; the grader is the institution to read, not the model.

Score basis. Atk 6 is reach: Anthropic’s Opus 4 card says the uplift trial was “graded by Deloitte using a detailed rubric,” that Deloitte’s red team “noted substantially increased risk in certain parts of the bioweapons acquisition pathway” for Opus 4, and ties both to the ASL-3 decision; 15 cards name it, Meta’s and Amazon’s reports name it, and the UK MoD gave Deloitte LLP the AI Assurance lot (Anthropic, Opus 4; Meta, Muse Spark report; Amazon, Nova report; Find a Tender). The points under 7 are that no statute or gate requires a lab to ask, that the work is one domain, that OpenAI’s pages name Gryphon and not Deloitte, and that Anthropic’s three newest cards drop it; SecureBio at 7/7/7, hazard 72, is named by four labs and sits above it, and HiddenLayer at 6/4/5, hazard 50, holds the same reach and entrenchment with a point less insulation. Def 5 is funding insulation: FY2026 revenue of US$74.5 billion means no lab is a large share, and no lab states a fee, but Anthropic’s announcement quotes Deloitte Global on “this significant investment in Anthropic’s AI platform” and Meta’s CEO names Deloitte as a Llama partner (Deloitte, FY2026 revenue; Anthropic, 6 October 2025; Meta, Open Source AI); Faculty at 8/5/6, hazard 63, holds the same insulation as a client-funded firm owned by Accenture, and Scale AI at 8/6/8, hazard 81, one point more. Hp 5 is entrenchment: nineteen months in Anthropic’s cards and a UK award from December 2025, against three newer cards that do not name it and no statute or program of record behind it (Anthropic, Opus 5.5); Trajectory Labs at 7/4/5, hazard 53, holds the same entrenchment and the same hazard on a different mix. Hazard 53 follows the (atk+def+hp) x 10/3 convention its peers follow. The values hold.

Sources: Deloitte US, Deloitte Expands End-to-End AI Controls and Assurance Capabilities; Deloitte UK, AI Assurance; Deloitte UK, Trustworthy AI; Deloitte US, Deloitte and Anthropic Alliance; Deloitte, Deloitte Acquires Gryphon Scientific; Deloitte, Deloitte reports FY2026 revenue; Deloitte will make Claude available to 470,000 people: Anthropic; Open Source AI Is the Path Forward: Meta; The Amazon Nova Family of Models: arXiv; Muse Spark Safety & Preparedness Report: Meta, arXiv; Claude Opus 4 and Sonnet 4 system card: Anthropic; Claude Opus 5.5 system card: Anthropic; Claude Haiku 5.5 system card: Anthropic; GPT-6 Astra deployment safety: OpenAI; AI Technical & Data Science Support, contract award notice: Find a Tender; USAspending: NIH award 75N97A26F10001; AIUC-1 consortium; Georgetown Journal of National Security Law and Policy.