LOGAN GRAHAM

Behavioral Archetype

THE ADVISER WHO BECAME THE EVALUATOR — Subject spent two years as the Prime Minister’s lead adviser on artificial intelligence, technology security and data policy, left Downing Street on 8 July 2022 under a two-year bar on lobbying the UK Government for his next employer, and joined Anthropic on 15 November 2022 as a machine learning research engineer. He now heads the team that tests the company’s models for the capabilities governments worry about, and he is the person the company sends to Congress to explain what those tests found. This profile scores the seat, an evaluation function whose findings reach governments before they reach the public, and not the man.

Documented conduct. The pattern is the adviser’s: find the finding, then brief the people who can act on it. His statement for the 17 December 2025 House hearing opens with the company’s valuation and its forecast, “by late 2026 or early 2027, it may be possible to have ‘a country of geniuses in a data center,’” then reports a measured trend (“this year, they have begun outcompeting humans in some head to head competitions”) and ends with four asks of Congress: codify CAISI, build threat-intelligence channels, adopt transparency rules and keep the chip export controls (Statement for the Record, docs.house.gov). His own website gives the same method in one line: “Our mission is to discover the most advanced (and possibly risky) capabilities of AGI, as far ahead as possible” (logangraham.xyz).

Essence Indicators

• Head of the Frontier Red Team at Anthropic since November 2022 by his own CV, which says he “built and directed a team of 15 researchers and dozens of external contractors” and “designed, directed, and built the first system to test whether the most advanced AI models may present national security risks” (CV filed with the House, docs.house.gov). The House committee lists him as “Dr. Logan Graham, Department Head, Frontier Red Team, Anthropic PBC” (House Homeland Security, advisory of 15 Dec 2025).

• Before Anthropic: Special Adviser at 10 Downing Street. The UK Government’s Annual Report on Special Advisers lists “Logan Graham” at pay band 3, £80,000-£84,999, among the special advisers in post on 30 June 2022 under the Prime Minister (Cabinet Office, Annual Report on Special Advisers 2022). His CV gives the post as “Special Adviser to the Prime Minister, July 2020 - July 2022” and the portfolio as “Lead adviser to the Prime Minister on artificial intelligence” and on R&D funding, technology security and data policy.
• The exit is on the Cabinet Office’s business-appointment record: applicant Logan Graham, former role “Special Adviser,” left 8 July 2022, new employer Anthropic, “Machine Learning Research Engineer (15th November 2022),” with a two-year bar on drawing on privileged information, on “directly involved in lobbying the UK Government on behalf of Anthropic,” and on advising on bids for Government funding or contracts (Cabinet Office, business appointment rules advice, Oct to Dec 2022).

• By his own account he did more at Downing Street than advise: his CV says he “conceived of and wrote the UK National AI Strategy, started the $1bn compute review, established the UK’s AI talent recruitment program,” and his website says “We made AI a national priority, created ARIA, allocated billions to science, made the Prime Minister’s office a startup, and worked to prevent the largest national security risks” (logangraham.xyz). These are his claims about his own work; the authorship of government documents is the government’s to confirm.
• Education: D.Phil. in Engineering Science at Oxford (2015-2020, “Recipient of the 2015 Rhodes Scholarship”), B.A. in Economics at the University of British Columbia, per the same CV; also Babylon Health (research scientist, 2019), X (AI resident, 2019-2020), and fellowships at Interact and South Park Commons (2021-2023).
• The hearing (17 December 2025). A joint hearing of the House Homeland Security Subcommittees on Cybersecurity and Infrastructure Protection (Chairman Andy Ogles) and on Oversight, Investigations, and Accountability (Chairman Josh Brecheen), “The Quantum, AI, and Cloud Landscape: Examining Opportunities, Vulnerabilities, and the Future of Cybersecurity,” 10:00 a.m., 310 Cannon. The Anthropic witness sat with Royal Hansen of Google, Eddy Zervigon of Quantum Xchange and Michael Coates of Seven Hill Ventures (House Homeland Security advisory). Livestream: YouTube.
• What he told the committee about the case. “Our preliminary estimate is that the threat actor was able to leverage Claude to perform the work of a 10-person team managed by one human operator”; “approximately 80 to 90% of the CCP-backed campaign tasks were automated by Claude, whereas the remaining 10 to 20% were tasks where the human operators reviewed Claude’s outputs and directed the models.” The limits he named: “the models frequently hallucinated” and “the attack still fundamentally required a human operator at various decision points.” The attribution rests on “technical evidence from the sophisticated obfuscation infrastructure,” targets that “aligned with known targets of the CCP,” and operators “following the Chinese workday – including observing lunch breaks – and observing Chinese national holidays” (Statement for the Record). Anthropic’s own post of 13 November 2025 gives the headline figure as “80-90% of the campaign” and “roughly thirty global targets” and names no individual (Anthropic).

• What he asked Congress for. On pre-deployment testing: “Codifying the CAISI can ensure the government can test and evaluate models for these capabilities, in partnership with the U.S. national security community,” and “The Frontier Red Team has an ongoing partnership with the U.S. government that enables risk mitigation and provides strategic national security insights.” On information sharing: “there should be more robust channels between American frontier AI laboratories and the U.S. government to facilitate threat intelligence sharing, similar to information sharing processes used in critical infrastructure sectors.” On transparency: Anthropic “published a transparency framework to inform light-touch guardrails that encourage the largest AI developers to follow secure practices.” On chips: “Relaxing controls on advanced AI chips at this juncture could allow the CCP to close the gap in frontier AI development” (Statement for the Record). See US CAISI.
• The disclosure form. His Truth in Testimony form lists no federal grants or contracts on the subject, and says “Anthropic maintains a voluntary MOU with the U.S. CAISI at the Department of Commerce for pre-deployment testing of our AI models; this is a non-compensatory partnership,” and a “voluntary collaboration agreement with the UK AI Security Institute for pre-deployment testing of our AI models; this is a non-compensatory partnership.” He lists himself as a fiduciary of no organization with an interest in the hearing (Truth in Testimony form, docs.house.gov).
• The research record he signs. Co-author of “Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims” (15 April 2020, 59 authors, listed 37th, with Miles Brundage, Jack Clark, Amanda Askell, Gillian Hadfield, Jade Leung and Markus Anderljung); of “Sleeper Agents” (January 2024, with Michael Sellitto and Jared Kaplan); and of “Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming” (31 January 2025, 43 authors) (arXiv:2004.07213; arXiv:2401.05566; arXiv:2501.18837).
• The team’s output. Anthropic’s Frontier Red Team page says the team “stress-tests AI systems to understand the full extent of their current capabilities and anticipate what comes next” and lists 34 posts, among them “Claude plays robotics” (9 Jul 2026), “Project Pilot: Can AI control a drone?” (24 Jul 2026), “Measuring tactical intelligence targeting and conventional weapons capabilities of AI models” (10 Sep 2026), “GLM-5.3 and the spread of advanced cyber capabilities” (29 Sep 2026), “AI models may be capable of self-replicating without access to a datacenter” (12 Oct 2026) and “Launching an opt-in vulnerability-finding service for open-source software” (8 Oct 2026) (Anthropic, Frontier Red Team). The page names no author for any post.
• June 2026. Simon Willison’s note on an Axios report of 15 June 2026 names “Logan Graham (I lead the Frontier Red Team at Anthropic), Dave Orr (Head of Safeguards…), and blog favorite Nicholas Carlini” as “reported to be meeting with the Commerce Department today in D.C.” (Simon Willison, 15 Jun 2026). The Axios page itself blocked the fetch; the line is attributed to Willison’s reading of it.

• His own stated beliefs. His website lists among its “core beliefs”: “Prepare for AGI,” “POSIWID” (the purpose of a system is what it does), “Accelerating science »> almost everything else,” and “There are no ‘adults in the room’; you are the adults; there is no room; this is the state of the world.” (logangraham.xyz).
• Federal giving. The FEC’s itemized-receipts file returns no contribution under “GRAHAM, LOGAN” with an Anthropic employer in the 2022, 2024 and 2026 cycles; the three small rows that match the name belong to a retired contributor in Augusta (FEC schedule A).

Social Persona / Impression Management

Immediate impression: A policy-adjacent engineer: Rhodes scholar, machine-learning doctorate, former Downing Street insider. The register on his own site is a start-up founder’s, not a civil servant’s: “We try to make our models do magic.”

Energy: Urgent and forward-leaning. The statement frames every finding as a date: “now is the moment to act,” two years ago, last year, this year, late 2026. He is the briefer, not the author of the headline; Anthropic’s own post carries the espionage numbers and names no one, and the testimony is where his name attaches to them.

Impression management strategy: The adviser’s method: put the finding in the form a minister can use, with a recommended action, and tie it to the national interest. The testimony’s frame is a race against China in which Anthropic’s detection capability exists because of US chip policy: “Because CCP-sponsored operators had to use our systems, we were able to detect and disrupt them, and share information about the threat with the U.S. government. That is an enormous strategic advantage.” The same sentence makes the company’s product the sensor and the export control the reason it works (Statement for the Record).

• Social accounts and feeds. X @logangraham, LinkedIn logangraham and GitHub, each linked from his own site and CV (logangraham.xyz; CV). No personal feed or blog feed; his site is a single page. The team’s publications are on Anthropic’s Frontier Red Team page, which lists posts by date and names no author. LinkedIn is not ground truth.

Forensic Archetype Comparison

Psychometric Assessment

Big Five (OCEAN):

Dark Triad:

MBTI: INTJ — Dominant introverted intuition, auxiliary extraverted thinking. The statement is a forecast with a plan attached. The score is an observation of the documented public role, from the exhibits cited in each row, not a claim about private character.

Threat Assessment

Score basis (no prior value). Hazard 50, components 6/4/5 (the formula gives 50.0). Leverage 6: the statement says his team evaluates cybersecurity, biosecurity and autonomy for a company valued by its own footnote at $183 billion (WSJ, 2 Sep 2025), and that government testers see the models first, which makes his function an input to release rather than the decision (Statement for the Record). Not higher: no gate bears his name. Below Jacob Klein (8/7/7), whose function bans accounts at the rate of 1.45 million a half-year. Insulation 4: an employee, no ownership, partnerships that are “voluntary” and “non-compensatory” in his own words (Truth in Testimony form). Durability 5: the team’s archive of 34 posts and the co-authored papers persist beyond any one head (Anthropic).

Alignment Analysis

Stated alignment: Find the dangerous capabilities first and help defenders: “Anthropic is determined to support defenders,” and the team’s mission is “to accelerate coordination around AI safety by producing evidence of emerging risks” (Statement for the Record; Anthropic).

Observed alignment: The findings travel with recommended rules, and each rule matches a practice the employer has already adopted: a voluntary CAISI MOU, a published transparency framework, existing threat-sharing with other labs, and a stated position on chip controls (Statement for the Record).

Gap assessment: The stated and observed alignments coincide on the evidence and diverge on who writes the rule. The record cannot settle whether the asks follow from the findings or from the employer’s position, and for the seat it does not need to: the finding and the interest are both in the same statement, each labelled.

Interest: Documented, never motive, never a verdict.

• Position: “Codifying the CAISI.” What adoption pays: Anthropic already holds a voluntary MOU with CAISI and says the testing “allows us to launch our commercial models more rapidly and with enhanced confidence about their reliability,” so a statutory CAISI turns an arrangement the company already passes into a requirement on every developer; the cost lands on labs without the relationship (Statement for the Record). His best case: “the CAISI and the U.S. government in general are in an advantageous position to evaluate model capabilities and understand capability trajectories better than anyone in the world.” Others who gain: the Intelligence Community and the Department of Energy, whose “unique expertise” the statement names, and CAISI itself. See US CAISI.
• Position: lab-government threat-intelligence channels. What adoption pays: Anthropic is the party that detected the campaign and shared indicators “with partners, including other frontier labs, with whom we have threat sharing agreements”; a formal channel makes that practice the standard. His best case: “We know that all American frontier AI labs are targets for infiltration by state and non-state actors.” Others who gain: the other frontier labs and the agencies on the receiving end.
• Position: “light-touch” transparency for “the largest AI developers.” What adoption pays: the framework is one Anthropic published on 7 July 2025, so the rule mirrors its own practice and applies only to the largest developers (Anthropic, The Need for Transparency in Frontier AI). His best case: transparency “would establish industry best practices for safety and set a baseline for secure model training.” Others who gain: the public and regulators that would read the disclosures; the smaller developers the framework exempts.
• Position: keep export controls on advanced chips. What adoption pays: a US lead in frontier models is the company’s competitive position, and the testimony ties its detection of the CCP-sponsored group to that lead. His best case: “Export controls on advanced semiconductors have proven effective at constraining the CCP’s AI development. Without them, what any individual American company does to secure its own models becomes far less consequential.” Others who gain: the US-based model developers; the opposing case, from chip makers, is not in this record.

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong.

Alliance Map

Two of these run in parallel: the Agency rows record a man who advised the UK’s Prime Minister and now briefs the US Congress and Commerce Department on behalf of a company that tests its models with both countries’ AI institutes. The Cabinet Office bar on lobbying the UK Government for Anthropic ran to about July 2024; the House statement is the first public advocacy of record on the US side. The overlap is the finding, not a coordination claim.

Convergent Drive Classification

Self-preservation: His standing is attached to the employer’s: the statement opens with the company’s valuation and describes its detection of the campaign as its own success (Statement for the Record).

Goal preservation: The goal stated on his own page has not moved since he joined: “discover the most advanced (and possibly risky) capabilities of AGI, as far ahead as possible” (logangraham.xyz).

Resource acquisition: Team of 15 and “dozens of external contractors” by his CV; access to the government through two voluntary testing agreements (CV; Truth in Testimony form).

Self-improvement: Each stage widens the audience: a machine-learning thesis, the Prime Minister’s office, a lab red team, a House hearing. The team’s own page lists 34 posts.

Subject is not an AI system. The drives appear anyway — in the evaluator.

Sources: Statement for the Record, Dr. Logan Graham, 17 Dec 2025; Biography/CV filed with the House; Truth in Testimony disclosure form; House Homeland Security Committee advisory, 15 Dec 2025; Cabinet Office, Annual Report on Special Advisers 2022; Cabinet Office, business appointment rules advice, Oct to Dec 2022; Logan Graham, personal site; Anthropic, Disrupting the first reported AI-orchestrated cyber espionage campaign; Anthropic, Frontier Red Team; Anthropic, The Need for Transparency in Frontier AI; arXiv:2004.07213; arXiv:2401.05566; arXiv:2501.18837; Simon Willison, 15 Jun 2026; FEC schedule A.