APART RESEARCH

Institutional Archetype

THE TALENT FUNNEL WITH AN EVALUATIONS DESK — Apart Research runs weekend research hackathons it calls Sprints, takes the best projects into a fellowship, and publishes evaluation work of its own. Its About page says it “develop[s] research and evaluations focused on harmful manipulation by advanced AI” and that “Through Sprints and Fellowships, we also help researchers, engineers, and domain experts turn promising ideas into rigorous research” (Apart Research, About).

Most evaluators on this site test models. Apart also supplies the people who go on to test them: its fellowship page says “Fellows have joined organizations like Anthropic and METR or founded startups like Geodesic Research and Asymmetric Security” (Apart Research, Apart Fellowship). One organization is a recruiting channel for the field and a small evaluator at once. Its board includes the CEO of Redwood Research, and its co-organizers include Redwood, METR and the UK AI Security Institute.

Mandate & Origin

“Apart Research is a nonprofit AI safety organization.” The About page headline is “Accelerating AI safety through research, evaluations, and field-building”; its footer says “Apart Research advances AI safety through rapid research, rigorous evaluations, and programs that bring new talent into the field.” Its home page says “We study and evaluate manipulation risks in frontier AI systems. We run applied research programs to bring global talent to AI safety.” (Apart Research, About; Apart Research).

The chief executive is Jaime Raldúa Veuthey; the About page lists Esben Kran under “Board and advisors” as a Board Member, alongside Buck Shlegeris, Mathias Kirk Bonde and Finn Metz, with Jason Hoelscher-Obermaier as Research Advisor and Anna Wisakanto as Strategy Advisor. The same page describes Raldúa’s earlier work: “Before Apart, Jaime did robotics research, worked in software engineering, and founded a consultancy helping EA organizations with tech and data.” Kran founded the organization: in a July 2023 Manifund application he is listed as the project’s creator, the application names “Esben Kran (CEO) and Fazl Barez (CTO),” says “Esben founded Apart” and describes the Alignment Jam hackathons as “initiated in November 2022”; in 2024 and 2025 Apart’s newsletters carry his byline as “Co-Director” beside the Research Director Jason Schreiber (Apart Research, About; Manifund; Apart Research, Code Red wrap-up; Apart Research, newsletter).

The legal entity is Apart Research Inc, 1500 N Grant St, Denver, Colorado. ProPublica’s Nonprofit Explorer lists it as a 501(c)(3) under EIN 41-5214746 with “Tax exemption issued: May 2026” and no Form 990 data, so the filing record that other evaluators on this site carry does not yet exist for Apart (Apart Research, About; ProPublica Nonprofit Explorer). Before that date Apart’s 2024 grant from the Survival and Flourishing Fund was routed through a different recipient entity, Ashgro Inc. (SFF 2024).

Funding & Backers

The About page and home page name no funders, and the media kit says only “Apart Research is a nonprofit working on independent AI safety research and evaluations” (Apart Research, Media kit). Its funders are disclosed by the funders and in Apart’s own fundraising newsletters. See The EA to AI-Safety Funding Pipeline.

• Coefficient Giving (formerly Open Philanthropy), from its own grant feed. “Open Philanthropy recommended two grants totaling $130,050 to Apart Research to support their work hosting four “hackathons” where participants will work on small projects related to AI alignment,” posted 3 April 2023 under potential risks from advanced artificial intelligence. “Open Philanthropy recommended a grant of $161,280 to Apart Research to support its work hosting monthly “hackathons” where participants will work on small projects related to AI alignment,” posted 3 June 2025, which “follows our December 2022 support” and falls under Global Catastrophic Risks Capacity Building. The two postings come to $291,330 (Coefficient Giving, grants search). See Open Philanthropy.
• Survival and Flourishing Fund. SFF-2024: Apart Research, $294,000 in the Main track, recipient entity Ashgro Inc., “General support of Apart Research,” Funder Jaan Tallinn. Apart’s July 2023 Manifund application said “We are receiving a speculation grant from SFF that supports Apart’s operations, partly including the hackathons.” The 2025 and 2026 recommendation tables do not list Apart (SFF 2024; SFF 2025; SFF 2026; Manifund). See Jaan Tallinn.
• The 2025 public fundraiser. Apart’s 7 June 2025 newsletter says “our current funding expires soon, and without immediate support, our momentum may stop,” and explains: “A foundation we expected to support our work because we outperformed last year’s grant significantly was not accepted.” On 17 June it reported “another $462,276 since our last newsletter, making the total $597k of our $955k goal,” and on 11 July: “In total, we received donations from more than 70 private donors along with support from Jaan Tallinn’s foundation and AISTOF.” That closing post set “Our next goal is $954k, which will secure sustainable funding for 12 months as we make room to grow with institutional funders.” (Apart Research, 7 June 2025; Apart Research, 17 June 2025; Apart Research, 11 July 2025).
• Sponsors of individual programs. Sprint tracks and prizes carry named sponsors: Equistamp on the 2026 AI Control hackathon, CBAI, Measuring AI Progress, Fourth Eon Bio and Sentinel Bio on the 2026 AIxBio hackathon, Halcyon Futures on the November 2025 Def/acc hackathon, Modal and the AI Alignment Foundation on the October 2026 Collusion sprint. The first Partnered Fellowship, with Martian, “secured over $70,000 in research support for the fellows” (Apart Research, AI Control 2026; Apart Research, AIxBio 2026; Apart Research, Def/acc 2025; Apart Research, Collusion Sprint; Apart Research, Partnered Fellowships).

Institutional Voice & Intent

Apart’s own account of what it sells, from its home page, addressed to three audiences: to “Frontier AI labs,” “Targeted model-behavior evaluations, independent testing of safeguards, and evidence for internal safety decisions”; to policymakers, “Empirical studies tied to policy questions, reproducible methods, and technical reporting” (Apart Research). On the sprint model: “Research hackathons, about once a month, online and at local sites. Teams build something others can check: an experiment, an evaluation, a dataset or a tool” (Apart Research, Sprints).

Stated intent: Produce independent evidence about harmful manipulation and other risks, and widen the pool of people who can do that work.

Observed intent: Run a pipeline that starts with a weekend event, selects “usually the top 5 to 20%” into the Apart Studio, takes “about 40% of Studio teams” into the fellowship, and places fellows in the organizations that run evaluations, while building a Research Division whose first subject is manipulation and whose first named audience is the EU AI Office (Apart Research, Apart Fellowship; Apart Research, Research).

Gap: Apart discloses its program selection rates, its publication list and the sponsors of individual events. What its pages leave out is a funder list, a client list for the evaluations offered to labs, and any evaluation result for a named lab model; its research page lists benchmark and method papers and a Science policy forum article, and no lab names it as a tester in a document cited here. The record shows an organization that is a recruiter and a policy contributor first and an evaluator second. For a pipeline, that is the stated design.

Interest: Apart argues for independent, proportionate, empirical evaluation of manipulation risk and for open entry to AI safety research. What adoption would pay it is set out from its own pages; no fee or contract value is published.

• Independent evaluation of harmful manipulation under the EU AI Act. Position, from its research page: “This work supports the EU AI Office as it prepares to enforce the EU AI Act. Our involvement spans research published in Science, the Code of Practice, the AI Act Advisory Forum, the Expert Forum on Frontier AI, a harmful-manipulation evaluation tender and much more.” What it pays: a Research Division with open roles for a “Research Scientist, Manipulation Evals” and a media kit that says “Apart tests general-purpose AI models for harmful manipulation, one of the systemic risks under the EU AI Act”; an evaluation requirement for that risk class needs evaluators, and Apart is positioned as one (Apart Research, Research; Apart Research, Media kit; Apart Research, Careers). Best case, from the summary line on its research page for its Science policy forum paper: “AI evaluations should provide meaningful risk information without imposing excessive burden.” Convergence: FAR.AI says it will support “Harmful Manipulation Risk Modelling Evaluation” as a subcontractor on Lot 4 of an AI Office tender, so a second evaluator on this site works the same buyer’s manipulation question; neither Apart’s page nor FAR.AI’s names the other or the lot’s lead (FAR.AI blog).
• Open entry to the field through hackathons and fellowships. Position: “One weekend Sprint has been enough to start many participants on research careers.” What it pays: sponsors and partners buy access to the pool; its fellowship page invites partners to “Run a Partnered Fellowship on your research agenda, or sponsor one. We handle recruiting, project management and compute,” and its first partner, Martian, set the research agenda for a fellowship (Apart Research, Apart Fellowship; Apart Research, Partnered Fellowships). Best case, its own: “Partners set the agenda and mentor the teams; Apart runs the program, with project management, compute and publication support.” Convergence: the organizations Apart’s fellows join, among them Anthropic and METR, recruit from a pipeline their own funders pay for; Coefficient Giving classifies Apart’s grant under Global Catastrophic Risks Capacity Building (Coefficient Giving, grants search).

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.

Position in the Apparatus

• Sprints. The Sprints page offers “See all 59 Sprints,” the next of them the AI Collusion Research Sprint, 23 to 25 October 2026; the About page counts 50+ Sprints run, 8,700+ Sprint participants, 1,600+ Sprint projects submitted, 60+ publications in AI safety, 100+ researchers in the Apart Fellowship and 120+ nationalities in its programs (Apart Research, Sprints; Apart Research, About).
• Benchmarks and methods. DarkBench, “A benchmark of 660 prompts examining manipulation, sycophancy and other dark patterns in leading language models” (ICLR 2025, oral); a noise-injection method for detecting sandbagging (NeurIPS 2025); the Catastrophic Cyber Capabilities Benchmark (2024); and “Multi-Agent Risks from Advanced AI,” a Cooperative AI Foundation technical report (2025) (Apart Research, Research).
• Policy. “The science and practice of proportionality in AI risk evaluations,” a Science Policy Forum piece published 19 February 2026, has 22 authors, of whom Jaime Raldua, with an Apart Research affiliation, is one; the others are drawn from the European AI Office, the Joint Research Centre, RAND, Safer AI, Johns Hopkins, CISPA, the Alan Turing Institute, METR and Irregular, among others (Crossref record, doi:10.1126/science.aea3835; Apart Research, Research).
• The METR tasks. In 2024 Apart ran the Code Red hackathon with METR: 128 participants, 231 task ideas and 28 full implementations following METR’s task standard, with “$30,000 in prizes” judged by METR researchers. Apart wrote “Some of these tasks are currently in use by the UK AI safety institute.” METR’s statement, quoted by Apart: “Collaborating with Apart for the Code Red Hackathon has been useful to METR’s work on developing a set of evaluation tasks” (Apart Research, Code Red wrap-up). See METR.
• AI control. The March 2025 AI Control hackathon was “co-organized by Redwood Research, and the UK AI Safety Institute” as an extension of ControlConf; the March 2026 edition “is the second edition of the hackathon, co-organized with Redwood Research (the group that founded the field of AI control) and sponsored by Equistamp” (Apart Research, AI Control 2025; Apart Research, AI Control 2026). See Redwood Research.
• Co-organized Sprints, 2025 to 2026. Economics of Transformative AI with BlueDot Impact (April 2025); AI Incident Response with CeSIA (September 2026, more than 230 projects); Digital Minds with the NYU Center for Mind, Ethics & Policy, Eleos AI Research and the California Institute for Machine Consciousness (August 2026, 237 projects); Secret Loyalties with Formation Research (July 2026); Secure Program Synthesis with Atlas Computing (May 2026, 61 projects), which fed a four-month fellowship; Swarm Alignment with the AI Alignment Foundation in partnership with AE Studio (November 2026); and AI Collusion with Poseidon Research and AE Studio (October 2026) (Apart Research, Sprints and the Sprint pages cited in the Alliance Map).
• Partnered Fellowships. The Martian Fellowship on Mechanistic Interpretability (September to December 2025), the Heron AI Security Fellowship (January to April 2026) and the Secure Program Synthesis Fellowship (June to September 2026), each listed as finished (Apart Research, Partnered Fellowships).
• People who came from the evaluators. Research Scientist Jerome Wynne “has developed evaluations and research tools at the UK AI Security Institute and Prolific AI Research”; Research Scientist Lorenzo Pacchiardi advised “the European Commission in supporting the implementation and enforcement of the EU AI Act in various capacities, most recently as a member of the AI Act Scientific Panel”; Evaluations Engineer Jaco Du Toit built “agentic evals for the UK AI Security Institute” (Apart Research, About).

Actions & Leadership Choices

CONDUCT: WATCH — PIPELINE AND EVALUATOR IN ONE ORGANIZATION. No wrongdoing is alleged. The structural point is overlap: the organization that recruits and trains evaluation talent also sells evaluations, holds a board seat for the CEO of another evaluator, co-hosts events with that evaluator and with a government institute, and says it supports the EU AI Office’s work on a risk class that evaluators are paid to test. None of it is hidden. It is how small the field is.

• November 2022. The Alignment Jam hackathons begin (Manifund).
• June 2024. Code Red hackathon with METR wraps; tasks are in use by the UK AI Safety Institute (Apart Research, Code Red wrap-up).
• 3 June 2025. Coefficient Giving posts a $161,280 grant for monthly hackathons (Coefficient Giving, grants search).
• 7 June to 13 July 2025. Public fundraiser after a foundation declines its application (Apart Research, 7 June 2025; Apart Research, 11 July 2025).
• 3 December 2025. Introduces Partnered Fellowships (Apart Research, News).
• 19 February 2026. Co-authored Science policy forum article on proportionality in AI risk evaluations (Crossref record).
• May 2026. IRS exemption issued to Apart Research Inc (ProPublica Nonprofit Explorer).

Alliance Map

Where ties on different axes run at the same time, the overlap is the finding. Apart’s funders are the network that funds FAR.AI, Palisade Research and Epoch AI: Coefficient Giving appears in the funder records of all four, and SFF in those of FAR.AI and Palisade Research. Its board seats the head of Redwood Research, which co-organizes its AI control events; its fellows go to METR and Anthropic, and METR’s evaluation tasks came from its hackathon; its page names a harmful-manipulation evaluation tender at the EU AI Office, and FAR.AI’s names Lot 4 of an AI Office tender for the same subject; neither page ties the two. The documents show disclosed arrangements, not coordination.

Reach Assessment

Institutional: Talent supply to the evaluators and a policy channel into the EU AI Office. Memetic: DarkBench, an ICLR 2025 oral with its own benchmark site, is the work most likely to travel outside the field.

Civilizational: If an organization that runs the funnel into AI safety work and also sells evaluations to the labs and the AI Office becomes the default place new evaluators learn their craft, the field’s methods inherit its curriculum: manipulation, collusion, control, digital minds, secret loyalties. Apart’s pages are open about the structure: selection rates, the sponsors of each track, and the partner who sets each fellowship’s agenda. What a reader cannot check is how much of its income comes from evaluation work, because it publishes no funder list, no client list and, until May 2026, no exempt-organization record.

Score basis. Atk 5 is reach: Apart’s research page says its work “supports the EU AI Office as it prepares to enforce the EU AI Act” and names the Code of Practice, the AI Act Advisory Forum and a harmful-manipulation tender; its About page counts 8,700+ Sprint participants; and its 2024 METR hackathon produced tasks it says the UK AI Safety Institute uses (Apart Research, Research; Apart Research, About; Apart Research, Code Red wrap-up). The points under a higher number are that no statute or release gate makes any lab ask for its tests and that no lab release document cited here names it. Peers on atk 5: Epoch AI, no statute either, with 987,000 site users and central banks citing it; Andon Labs holds 6 because three labs put its work in release documents; FAR.AI holds 7 on a lab system card, an EU consortium lead and a public leaderboard. Def 5 is funding insulation: more than 70 private donors, Jaan Tallinn’s foundation and AISTOF named in its own fundraiser post, and no lab payment in any record cited (Apart Research, 11 July 2025). The points off are that its own site lists no funders, that Coefficient Giving’s $291,330 and SFF’s $294,000 are one philanthropic network’s money, and that its June 2025 newsletter says its funding was about to expire (Coefficient Giving, grants search; SFF 2024; Apart Research, 7 June 2025). Peers on def 5: Andon Labs, whose funding is undisclosed and whose access depends on the labs it grades; Palisade Research and FAR.AI hold 6, Palisade on nearly all its money from the same two networks with disclosed lines, FAR.AI on a published cap of 10 percent of revenue from for-profit developers. Hp 5 is entrenchment: 59 Sprints, 100+ fellows, 60+ publications, Partnered Fellowships with outside organizations, and a 501(c)(3) exemption issued in May 2026 (Apart Research, Sprints; Apart Research, Partnered Fellowships; ProPublica Nonprofit Explorer). The points off are that no statute, contract or program of record installs it and that a lab or evaluator can recruit from another pipeline. Peers on hp 5: Palisade Research, FAR.AI and Andon Labs hold 6 on installed deployments, meetings series or filed accounts; Epoch AI holds 7 on installed datasets and a three-year donor renewal.

Sources: Apart Research — home; Apart Research — About; Apart Research — Research; Apart Research — Sprints; Apart Research — Apart Fellowship; Apart Research — Partnered Fellowships; Apart Research — Media kit; Apart Research — Careers; Apart Research — News; Apart Research — fundraiser update, 7 June 2025; Apart Research — fundraiser extended, 17 June 2025; Apart Research — final fundraiser newsletter, 11 July 2025; Apart Research — Code Red wrap-up; Apart Research — AI Control Hackathon 2025; Apart Research — AI Control Hackathon 2026; Apart Research — AI Collusion Research Sprint; Apart Research — Digital Minds Sprint; Apart Research — AI Incident Response Sprint; Apart Research — Secure Program Synthesis Hackathon; Apart Research — Secret Loyalties Hackathon; Apart Research — Swarm Alignment Hackathon; Apart Research — AIxBio Hackathon; Apart Research — Def/acc Hackathon; Apart Research — Economics of Transformative AI Sprint; Manifund — Run five international hackathons on AI safety research; Seldon Lab; Coefficient Giving — grants search; SFF 2024; SFF 2025; SFF 2026; ProPublica Nonprofit Explorer — Apart Research Inc; Science — The science and practice of proportionality in AI risk evaluations; FAR.AI — EU AI Act CBRN consortium post; Redwood Research — team.