Who's who in AGI: labs, safety orgs, and researchers

Why a personnel map matters, and why it expires quickly

Names and titles are not a substitute for understanding what a lab actually does — see how the frontier AI labs differ for that — but they matter anyway, because decisions about capability thresholds, deployment gates, and how much outside scrutiny a lab tolerates are made by specific people who can be replaced, promoted, or overruled. This page is a roster: who currently runs the frontier labs, which independent organizations do the outside checking, and which named researchers and public intellectuals shape the debate loudest. Every entry is dated and sourced so a returning reader can tell what has changed.

Frontier labs: who runs them, and where they say the frontier is headed

Lab	Current leadership	Positioning
OpenAI	Sam Altman, CEO and co-founder; Jakub Pachocki, Chief Scientist	Shipped GPT-6 “Astra” in September 2026. Days later, Pachocki published an essay arguing no lab — including OpenAI — has solved alignment well enough to justify scaling at full speed. Altman has said OpenAI will not pursue an IPO in 2026, citing safety concerns.
Anthropic	Dario Amodei, CEO and co-founder; Jared Kaplan, co-founder and Chief Scientific Officer	Positions itself as the safety-focused frontier lab. Amodei’s September 2026 essay called for deliberately slowing capability gains industry-wide, proposing independent evaluators inside labs and eventual international limits on recursive self-improvement.
Google DeepMind	Demis Hassabis, Chairman of DeepMind and Chief Scientist of Alphabet (moved from CEO in August 2026); Koray Kavukcuoglu, SVP running DeepMind day-to-day, reporting to Sundar Pichai	Underwent a major leadership reshuffle in August 2026, days after longtime chief scientist Jeff Dean departed after 27 years to found an independent company that Google is now investing in.
Meta AI (Meta Superintelligence Labs)	Alexandr Wang, Chief AI Officer, co-leading MSL; Shengjia Zhao, Chief Scientist of MSL	Wang joined via a roughly $14 billion Meta/Scale AI arrangement in June 2025. A parallel “Applied AI Engineering” unit added in March 2026 diluted his sole authority. Former Chief AI Scientist Yann LeCun departed in November 2025 to found a rival “world models” startup — see the skeptic case against near-term AGI for his technical argument.
xAI	Elon Musk, founder and chairman; Michael Nicolls, President (since April 2026)	Positioned explicitly to catch up to Anthropic’s Claude. Dan Hendrycks, director of the Center for AI Safety, serves as xAI’s safety adviser at a symbolic $1 salary to manage the conflict of interest. Corporate structure is in flux — see the caveats below.
Microsoft AI	Mustafa Suleyman, EVP and CEO of Microsoft AI (co-founder of DeepMind and Inflection AI), reporting to Satya Nadella	Shipped seven in-house “MAI” models at Build 2026 after Microsoft’s OpenAI contract was amended to permit independent frontier research. Suleyman has predicted roughly 18 months to human-level performance on most white-collar professional tasks.
Safe Superintelligence Inc. (SSI)	Ilya Sutskever, CEO (took over after co-founder Daniel Gross departed to Meta in 2025); Daniel Levy, President	Valued at $30 billion in a March 2025 round; explicitly single-product-focused, building toward safe superintelligence with no near-term commercial product.
Thinking Machines Lab	Mira Murati, co-founder and CEO (former OpenAI CTO)	Founded February 2025; raised the largest seed round in history at the time ($2 billion at a $12 billion valuation, July 2025). Shipped a fine-tuning API (“Tinker”) and real-time “interaction models”; reportedly in talks pushing valuation toward $50 billion.

Three items belong together because they landed almost simultaneously and read as a single moment rather than three unrelated stories. On September 6, 2026, OpenAI’s own chief scientist published an essay arguing that no frontier lab has solved alignment well enough to justify its current pace — a striking admission from inside the company shipping the newest frontier model. Days later, Sam Altman said OpenAI will not go public in 2026, citing safety concerns as a reason, and Anthropic’s Dario Amodei published his own essay calling on the industry to deliberately slow capability gains. All three are dated within the same week this page went to press — see the caveats section below before treating any of it as settled industry consensus rather than a live, contested moment.

Google DeepMind’s leadership changed shape in early August 2026: Hassabis moved out of the CEO seat into a chairman-and-chief-scientist role spanning DeepMind and Alphabet, while Kavukcuoglu took over daily operations. That reshuffle happened within days of Jeff Dean’s departure after 27 years at Google to start an independent research company that Google itself is now backing — an unusual arrangement worth watching rather than treating as resolved.

Meta’s AI effort has been reshaped twice in under a year: first by Wang’s arrival via the Scale AI deal, then by a March 2026 restructuring that added a second engineering unit reporting outside his chain. xAI’s picture is complicated further by an unresolved corporate-structure question — see the caveats below — on top of Hendrycks’s dual role advising the lab he’s also supposed to help keep in check. For the published policies these eight organizations say govern their own risk-taking, see frontier AI labs’ safety frameworks, compared — stated policy and organizational stability are two different things; a framework survives a CEO’s departure only if it was built to.

Independent AI safety research organizations

Frontier labs write their own safety frameworks. The organizations below don’t run frontier models themselves; their job is evaluating, auditing, or advocating from outside that structure — with real variation in how independent “outside” actually is.

Org	What it does
MIRI (Machine Intelligence Research Institute)	Founded 2000, the oldest AI safety org. Nate Soares is President; Malo Bourgon is CEO; Eliezer Yudkowsky is Founder, Senior Research Fellow, and board chair. Shifted toward public communication and policy advocacy, most visibly through Yudkowsky and Soares’s 2025 book arguing the default outcome of building superhuman AI is loss of control.
Redwood Research	Buck Shlegeris, CEO. Pioneers “AI control” protocols — designing deployment so a misaligned model can’t cause catastrophe even if it is misaligned — and co-demonstrated empirical “alignment faking” in large language models. Source: 80,000 Hours interview.
METR (Model Evaluation and Threat Research)	Beth Barnes, Founder and CEO; Chris Painter, President; Hjalmar Wijk, Chief Scientist. Runs pre-deployment autonomous-capability evaluations of frontier models and publishes the widely cited “time horizon” capability metric — see AGI timeline forecasts for the metric’s own documented limits. Source: metr.org/about.
UK AI Security Institute	Rebranded from “AI Safety Institute” in February 2025, shifting focus toward cyberattack, cybercrime, and fraud risk rather than broader AI safety and ethics. Adam Beaumont, a former GCHQ chief AI officer, is interim director; Jade Leung is CTO and also serves as the UK Prime Minister’s AI adviser.
US Center for AI Standards and Innovation (CAISI)	Formerly the “US AI Safety Institute,” renamed by the Trump administration in June 2025 to drop “safety” from its title and reorient toward national-security testing and commercial AI evaluation. Paul Christiano briefly served as Head of Safety here before resigning in August 2026 — see below. Source: nist.gov/caisi.
Apollo Research	Marius Hobbhahn, CEO and co-founder; Chris Akin, COO; Charlotte Stix, Chief Governance Officer. Specializes in detecting scheming and deceptive alignment — published, with OpenAI, the first evidence of in-context scheming in frontier models, and later helped develop a “deliberative alignment” technique that reduced measured scheming roughly 30-fold. Source: apolloresearch.ai/about.
GovAI (Centre for the Governance of AI)	Ben Garfinkel, Director since 2021, when it became an independent nonprofit under founding director Allan Dafoe. Research and a well-known fellowship pipeline aimed at helping government and industry navigate advanced AI. Source: governance.ai/about-us.
Center for AI Safety (CAIS)	Dan Hendrycks, Executive Director and Founder (also xAI’s paid-$1 safety adviser). Authored the influential May 2023 one-sentence statement holding that mitigating AI extinction risk should be a global priority alongside pandemics and nuclear war.
LawZero	Founded June 2025 by Yoshua Bengio, incubated at Mila. Builds deliberately “non-agentic” AI — a “Scientist AI” architecture with no hidden agenda — as a safety-by-default alternative to agentic frontier models. Source: Wikipedia.

The most consequential fact in this table isn’t any single org’s mission statement — it’s how much government “safety” institutes have been renamed and repositioned in the last eighteen months. The UK’s institute dropped “safety” from its name in February 2025 for cybercrime and fraud; the US institute did the same in June 2025, under a different administration, for different stated reasons. Neither change was cosmetic — both came with a stated shift in what the institute evaluates. That leaves genuinely independent evaluation work concentrated in a smaller set of nonprofits — METR, Apollo Research, Redwood Research, GovAI — whose funding and access still depend, in practice, on the labs choosing to cooperate.

Paul Christiano’s past year illustrates how thin the line between “government evaluator” and “industry insider” can get: from running the independent Alignment Research Center, to a part-time safety role inside CAISI, back to ARC as executive director in August 2026, and onto OpenAI’s own nonprofit-board Safety and Security Committee weeks later. None of that is improper on its face, but it means a small number of credentialed people frequently rotate between the organizations meant to check a lab and the lab’s own governance.

Named researchers and public intellectuals

Beyond the organizations above, a handful of individuals shape the AGI-risk conversation largely on their own authority — through public statements, technical work, or both.

Researcher	Current role	Known for
Yoshua Bengio	Full Professor, Université de Montréal; Founder and Scientific Advisor, Mila; Co-President and Scientific Director, LawZero	Chairs the International AI Safety Report (2026 edition describes accelerating capability alongside growing risk). Founded LawZero as a non-agentic alternative to frontier-lab architecture.
Stuart Russell	Professor, UC Berkeley (Smith-Zadeh Chair, EECS)	Leads the Center for Human-Compatible AI (CHAI), whose mission is “provably beneficial” AI via inverse reinforcement learning and related technical approaches. Elected to the National Academy of Engineering in February 2025.
Paul Christiano	Executive Director, Alignment Research Center (ARC)	White-box mechanistic-interpretability techniques for detecting misalignment. Rotated through a government safety role and onto OpenAI’s own board oversight committee within the same year — see above.
Geoffrey Hinton	No formal institutional safety role	2024 Nobel Physics laureate (with John Hopfield) for foundational AI work; the most prominent “godfather of AI” public voice on existential risk. Compared unregulated AI development to a car with no steering wheel at a 2026 conference in Geneva.
Eliezer Yudkowsky	Founder, Senior Research Fellow, and board chair, MIRI	Co-author, with Nate Soares, of the 2025 book arguing the default outcome of building superhuman AI is loss of control with existential consequences.
Dan Hendrycks	Executive Director and Founder, Center for AI Safety; xAI safety adviser	Authored the 2023 one-sentence extinction-risk statement (above). His dual role — running an independent safety nonprofit while advising a frontier lab — is itself a fact worth weighing when reading his public statements about either.

Two things are worth noticing about this list. First, it skews toward people with an institutional base in AI safety specifically — a deliberate scope choice; the separate case made by researchers who doubt near-term AGI is plausible at all (Yann LeCun, Andrew Ng, Gary Marcus, and others) gets its own dedicated treatment rather than folding in here. Second, several names above also appear in the tables higher up, in a different capacity: Bengio runs LawZero, Yudkowsky chairs MIRI’s board, Hendrycks founded CAIS and advises xAI. The overlap is real — the same handful of people wear multiple hats across the organizations meant to check the labs and the labs themselves.

What’s still moving as this goes to press

Four specific facts above are dated snapshots of situations that were actively changing at publication time, flagged here explicitly rather than presented as settled:

• Google DeepMind’s August 2026 reshuffle. Hassabis’s move to Chairman/Chief Scientist and Kavukcuoglu’s promotion happened within the six weeks before this page’s publication date. Titles and reporting lines at this altitude can shift again quickly; verify current leadership directly before relying on this row.
• The Amodei/Altman/Pachocki moment of September 6–13, 2026. Three statements from three labs’ leadership, landing the same week this page was written, argue in different ways that the industry needs to slow down or hasn’t solved alignment. It reads as a genuine inflection point, but it is also the most recent, least-settled material here — treat exact wording and full implications as provisional until more outlets have covered it independently.
• The UK AI Security Institute’s chair position. Adam Beaumont’s interim directorship is reasonably well documented; whether a separate chair role (previously held by Ian Hogarth) currently exists, and who holds it, could not be independently reconfirmed for 2026 in the sourcing behind this page. Confirm directly with the institute before citing a specific chair.
• xAI’s corporate structure. xAI has been described in 2026 reporting as folded into SpaceX ahead of a planned SpaceX IPO, with coverage referring to the combined entity informally as “SpaceXAI.” This is a fast-moving corporate-structure question rather than a settled fact; confirm the current legal and organizational relationship between xAI and SpaceX before treating it as fixed.

How to use this page

Treat this roster as a starting point for figuring out who is actually making a given decision, not as a verdict on any lab’s or researcher’s trustworthiness. Titles change, companies restructure, and the same person can hold a government role, a nonprofit role, and an industry advisory role within the same eighteen months — Paul Christiano’s path through CAISI, ARC, and OpenAI’s own board committee is one example above, and it won’t be the last. For what these labs actually commit to doing once a model crosses a capability threshold, see frontier AI labs’ safety frameworks, compared. For how fast the underlying capability itself is moving — and why forecasters disagree so sharply about it — see AGI timeline forecasts: what the labs and surveys say. And for the vocabulary this whole roster assumes, start with how the frontier AI labs differ.