Investigative thread alleges conflicts of interest across AI safety evaluation ecosystem tied to White House superintelligence pact

On October 17, @kevinnbass published a lengthy thread series (with visualizations and a public repo) mapping the funding and personnel ties across the AI model evaluation ecosystem, questioning the independence of the "independent external evaluators" named in the White House superintelligence agreement signed on September 29 by Google, Anthropic, Meta, OpenAI, xAI, and Nvidia. The author argues the evaluation ecosystem is deeply entangled with the labs it evaluates—financially, through investments, and through personnel—and that existing independence arrangements are insufficient to back the agreement's promises.

Confirmed

• Funding and investment networks: Backers such as Coefficient Giving (formerly Open Philanthropy, the Moskovitz foundation) span the evaluation ecosystem; of 43 organizations, 16 have records of funding from Coefficient Giving or Good Ventures, 16 have funding tied to Jaan Tallinn or Survival and Flourishing, 21 have a "triple conflict of interest" with Anthropic, and 39 of the 43 have documented ties.
• Investment overlap: 27 investors holding Anthropic shares also appear in at least 9 evaluation/benchmark-related institutions (Scale AI/SEAL, the Hugging Face alignment team, Apollo Research, etc.); Anthropic early backer Macroscopic Ventures also invested in Apollo Research (January 2026 seed round) and several other safety evaluation firms.
• Direct lab funding: In October 2023, Anthropic, Google, Microsoft, and OpenAI (founding members of the Frontier Model Forum) announced an AI Safety Fund exceeding $10 million, with recipients mostly evaluation organizations tied to the labs; 11 days before the agreement was signed (September 18), Anthropic announced "independent evaluations of frontier AI" led by Accenture-owned Faculty, with evaluators embedded inside Anthropic.
• Publication control: METR disclosed that in its May 2026 frontier risk report, every participating company exercised the right to redact or anonymize key conclusions before publication; METR's reports on GPT-5, GPT-5.1-Codex-Max, and GPT-5.6 Sol were marked "reviewed and approved at the request of OpenAI's communications and legal teams," and Anthropic's Claude Opus 5.5 system card likewise required review of METR reports.
• Commissioned evaluations and asset ownership: Epoch AI lists paid commissions including xAI's Grok 4 math evaluation and Google DeepMind's Gemini evaluation; OpenAI owns the majority of problems in the FrontierMath benchmark.
• Personnel and governance entanglement: Jason Matheny held an Anthropic trustee position while leading RAND, having previously served on the US National Security Commission on AI and in the White House Office of Science and Technology Policy; since April 2026, trustees appointed by the Anthropic Long-Term Benefit Trust hold a board majority, with the initial five trustees including RAND CEO Matheny; OpenAI safety committee chair Zico Kolter is also co-founder of external evaluator Gray Swan; when Open Philanthropy donated $30 million to OpenAI in 2017, it arranged for executive director Holden Karnofsky to join the board.
• Divergent payment policies: METR says it has not taken payments or donations from frontier AI companies to date (though it accepts substantial free model access), and in August announced roughly $71 million in pledges over six months without disclosing donors; Redwood Research says it did not charge OpenAI for one independent evaluation; Apollo Research charges market rates.
• Existing standards and author's self-check: In December 2025, METR, RAND, SecureBio, Transluce, and others formed the AI Evaluator Forum and published the AEF-1 standard setting minimum operating conditions for independent third-party evaluation; the author's self-audit found that of 8 unfavorable evaluations, only 3 were followed by new funding, and he concedes the sample is small and no causal claims are made.

Unconfirmed

• Whether funding "follows conclusions": The author admits the relevant sequence sample is small and not decisive—presented as observation, not causal conclusion.
• The other 4 organizations showed no ties found, but the author notes relevant records may not be public, which does not equal a clean bill of health.

Why it matters

• The White House superintelligence agreement makes independent external evaluation a key mechanism of its safety commitments; if evaluators have multiple funding, investment, and personnel ties to the labs they assess, the credibility of safety oversight faces systemic risk.
• Drawing on SEC independence rules from financial auditing, the author proposes reforms: evaluators and affiliated personnel should not hold direct or material indirect financial interests in evaluated companies, with mandatory disclosure of the "triple pack"—funding sources, investment relationships, and personnel movement—while noting the industry already has the codified AEF-1 standard ready to cite.

2026-10-17 ~ 2026-10-17 · 28 related posts

• [source] Who audits the auditors? Mega-thread maps financial ties between AI safety evaluators and the six White House Accord labs — kevinnbass · 2026-10-17
• 43-org map: 27 evaluator orgs tied to 4+ of 6 SI accord signatories, Anthropic most entangled — kevinnbass · 2026-10-17
• Map reveals 60 documented testing ties between 7 frontier AI labs and 26 evaluator orgs — kevinnbass · 2026-10-17
• 60 documented testing ties connect 7 labs to 26 evaluator organizations — kevinnbass · 2026-10-17
• METR raised ~$71M in commitments in six months without naming donors — kevinnbass · 2026-10-17
• METR reports on OpenAI and Anthropic models required lab review and approval — kevinnbass · 2026-10-17
• Funding map: 21 of 43 safety orgs backed by the same donors — kevinnbass · 2026-10-17
• Anthropic directly funds Accenture/Faculty for its 'independent' evaluations — kevinnbass · 2026-10-17
• [source] 39 of 43 organizations around Anthropic have documented ties — kevinnbass · 2026-10-17
• AI Evaluator Forum's AEF-1 sets minimum conditions for independent AI evaluations — kevinnbass · 2026-10-17
• OpenAI's board oversight tested: the 2023 Altman ouster, revisited — kevinnbass · 2026-10-17
• OpenAI safety committee chair Zico Kolter co-founded evaluator Gray Swan AI — kevinnbass · 2026-10-17
• Mapping 45 board seats: the intertwined safety governance of OpenAI and Anthropic — kevinnbass · 2026-10-17
• Government AI testers share personnel ties with Anthropic trust and OpenAI board — kevinnbass · 2026-10-17
• Did money follow findings? Of 8 adverse AI safety findings, 3 were followed by new funding — kevinnbass · 2026-10-17
• Thread proposes SEC-style auditor independence rules for AI safety evaluators — kevinnbass · 2026-10-17
• Epoch AI's paid evaluations for xAI and Google DeepMind; OpenAI owns most FrontierMath questions — kevinnbass · 2026-10-17
• Evaluator payment policies compared: METR takes no lab money, Apollo charges market rates — kevinnbass · 2026-10-17
• Labs' $10M AI Safety Fund grants went to many of their own evaluators — kevinnbass · 2026-10-17
• Macroscopic Ventures: Anthropic backer with stakes in Apollo Research, Gray Swan and Redwood — kevinnbass · 2026-10-17
• 27 Anthropic investors also documented in 9 evaluator organizations, data shows — kevinnbass · 2026-10-17
• Records trace Tallinn/SFF money conduits to AI evaluator EquiStamp — kevinnbass · 2026-10-17
• 18 of 43 AI safety evaluators lack conflict-of-interest policies — kevinnbass · 2026-10-17
• METR: AI labs redacted key claims before approving frontier risk report — kevinnbass · 2026-10-17
• [source] Mapping AI safety eval: 21 of 43 orgs tied to Anthropic via three or four channels — kevinnbass · 2026-10-17
• Jaan Tallinn/SFF funding map: 46 lines to 16 AI safety orgs — kevinnbass · 2026-10-17
• Mapping AI safety funding conduits: Anthropic, SFF, Coefficient's tangled ties — kevinnbass · 2026-10-17
• Mapping AI safety money conduits: Rethink Priorities, SAIF, Schmidt Sciences fund METR, Apollo, UK AISI — kevinnbass · 2026-10-17