DeepMind's 580 Pentagon Letter: The Most Sourceable AI Safety Pool

Fortune's August 10 exposé put a number on something recruiters at safety-first labs had only sensed: 580+ Google employees, including 20 directors and VPs, signed an open letter in April 2026 opposing the classified Pentagon Gemini contract. Demis Hassabis stepped aside as DeepMind CEO on August 5, becoming Alphabet chairman and chief scientist while retaining Isomorphic Labs. The London office voted to unionise in April with 98% in favour. If you hire on values fit, this is the largest self-declared candidate list the field has produced, and it has a shelf life.

Why the 580 signers are a better list than any skill filter

The Pentagon letter is a stronger sourcing artifact than a LinkedIn skill tag because it is a revealed-preference document: 580+ people accepted internal reputational risk to sign it, and their names are public. Skill filters catch aspiration. Public dissent catches conviction.

Here is the size mismatch that matters. In Refolk's index of professional profiles, only 265 people worldwide surface when you filter for "AI safety alignment" as a self-declared specialty. The April letter has more than twice that number of signatures from a single employer.

2.2x

That ratio is the whole thesis. If you are sourcing DeepMind engineers for a mission-aligned role, the signer list is not a supplement to your skill search. It is a superset of it.

The three concentric lists to work

Treat the public record as three tiers of decreasing size and increasing intent:

1. The 580+ Google-wide Pentagon letter signers. Includes Cloud, Search, and cross-functional staff. Broadest revealed preference. By early May one DeepMind researcher told Fortune the letter had reached roughly a thousand signatures.
2. The 100+ DeepMind-only internal weapons letter signers. Tighter, self-identifies as DeepMind research staff, correlates with the UK union-eligible population.
3. The 250+ co-signers of Alex Turner's 25-page human-oversight proposal. Named, self-declared, willing to co-sign a technical governance document that Google declined to implement. This is your priority tier.

A recruiter working the third list first, the second list second, and the first list as background will out-hire anyone starting from an "Ex-DeepMind + alignment" boolean.

The UK is the sourcing hotspot, not the Bay

London, not Mountain View, is where the DeepMind revolt has organisational form, which makes UK-based outreach materially faster than US outreach for the same pool. The union drive, the 98%-in-favour vote, and the senior-researcher energy behind the letter all originated in the London office.

The headcount split makes this counter-intuitive at first glance.

The US pool is 1.93x the UK pool by raw headcount. But the dissent-per-capita and the organising infrastructure are inverted. UK DeepMinders voted to join the Communication Workers Union and Unite the Union in April, and formally requested recognition. If granted, they would be the first unionised frontier AI lab workforce in the world.

For a recruiter, that means two practical advantages in London:

• A running attrition signal. Public CWU and Unite recognition-campaign communications double as a real-time morale ticker.
• Denser weak-tie networks. Union organising creates lists. Lists leak. A single warm intro in London gets you further than a hundred cold InMails in Mountain View.

If you are running AI safety researcher sourcing out of San Francisco, split the desk. Put a UK-based recruiter, or at minimum a UK-hours contractor, on the London 1,774. The Bay Area 3,425 needs a different, slower cadence.

The pool self-screens on compensation, so pitch differently

This cohort will filter comp-first outreach into the trash, because the act of public signing is a reputational commitment device that makes any move to a defense-adjacent employer visibly hypocritical. Alex Turner publicly declined outreach from the OpenAI safety team and, at time of writing, remains unemployed rather than take a role that fails his mission screen.

That is the mechanism you need to internalise. Signers cannot cash in without cost. So do not lead with cash.

What actually converts in this pool:

• A named, public safety mandate. Not "we care about safety." A specific policy: no defense contracts, publishable red-team results, independent review board.
• Publication rights. Turner co-authored Shard Theory and the Steering Vectors work behind Anthropic's Golden Gate Claude experiment. Researchers at that level need the paper trail.
• Governance detail. Turner's 25-page proposal asked for human oversight and independent reviews. If your startup can answer those exact asks in writing, you skip three screening rounds.
• Named leadership. Who is the CTO. Who is on the board. What did they publicly say about Project Maven, the April Pentagon deal, and autonomous targeting.

Signers cannot cash in without cost. Do not lead with cash. Lead with the governance document you would let them edit.

The pitch that works reads more like a policy memo than a job description. Which is the exact friction Refolk removes from the first step: you describe the person in plain English ("ex-DeepMind alignment researcher who publicly signed the April Pentagon letter, based in London, published on interpretability") and get a ranked shortlist without touching a boolean.

Anthropic is a bottleneck. Startups are the missing category.

The addressable-employer list for a values-consistent move out of DeepMind is exactly one large lab (Anthropic) plus a handful of academic and nonprofit labs, which means well-positioned startups have an unusually open lane. OpenAI, xAI, Nvidia, Microsoft, and Amazon have all signed similar Pentagon agreements. Only Anthropic refused.

The realistic destination map, as of August 2026:

• Frontier labs that pass the screen: Anthropic.
• Academic and nonprofit labs already indexed as top employers for global AI safety talent in Refolk's data: CHAI (Berkeley), Ai2, PIBBSS, BlueDot Impact.
• Everyone else: startups willing to write a defense-avoidance clause into their charter.

That is a bottleneck, and bottlenecks favour outreach speed. Noam Shazeer, Gemini co-lead, has already left for OpenAI. Nobel laureate John Jumper left for Anthropic. The senior exit lanes are visibly forming, and Anthropic's safety team can only absorb so many hires per quarter.

If you run a defense-averse startup, the pitch is: I am the third option. You do not have to wait in Anthropic's queue. You do not have to move to an academic salary. Here is my charter. That framing is the whole game for ex-DeepMind hiring in the second half of 2026.

The named seeds to work outward from

Start from the publicly named critics, then expand by co-authorship and co-signature, because the AI talent exodus of 2026 is graph-shaped: each named signer sits inside a cluster of co-authors, former labmates, and shared postdoc alumni. Two names do most of the graph work.

• Alex Turner. PhD in alignment from Oregon State. Postdoc at Stuart Russell's CHAI at UC Berkeley. Co-pioneered Shard Theory and Steering Vectors. Resigned over the Pentagon contract, mobilised the 250+ co-signers on the 25-page proposal, and turned down OpenAI safety-team outreach. His co-author graph is the densest single expansion in the pool.
• Andreas Kirsch. DeepMind research scientist. Publicly called the contract's restrictions "meaningless weasel words" in Fortune. His public posture is a beacon for like-minded colleagues who did not sign but agreed.

From these two, the expansions worth running:

1. Co-authors on their last three papers.
2. CHAI Berkeley alumni currently at DeepMind London.
3. Signers of the 25-page Turner proposal cross-referenced with the DeepMind-only weapons letter.
4. UK-based DeepMind employees who publicly liked or reposted CWU or Unite recognition communications.
5. Former Google Brain and DeepMind researchers already at Anthropic who could refer inward.

This is the sort of multi-source expansion that eats a sourcer's week if done by hand. Refolk collapses it into a single plain-English query across GitHub, LinkedIn, and the open web, which matters when your competitive window is measured against Anthropic's next hiring batch.

The window closes when Anthropic finishes its next hiring round

The pool is time-limited because Anthropic and a handful of nonprofits will absorb the most obvious names within one to two quarters, leaving behind a diffuse remainder that is harder to identify as values-first candidates. Once Turner takes a role, once the London union recognition is granted or refused, once Hassabis's successor is named, the coordinating signals fade. The signature list stays public, but the freshness of the signal decays.

Three dates worth watching:

• CWU and Unite recognition decision. Formal recognition or refusal will trigger a wave of either commitment or resignation.
• Anthropic's next safety-team hiring announcement. Anchors the "already placed" list.
• The next Fortune or NYT follow-up. Each round of coverage re-surfaces named critics and adds a few more.

Recruiters who treat the 580 as a static PDF will lose. Recruiters who treat it as a live graph, refreshed weekly against public signals, will pull the DeepMind-only 100 and the Turner-proposal 250 into offer stage before the story cools.

FAQ

How do I actually get the list of 580 signers?

The full signature list is not published in a single machine-readable file, but Fortune, TheNextWeb, and follow-on coverage name dozens of signers directly, and the internal DeepMind-only weapons letter and Turner's 25-page proposal have been referenced with partial signer counts and named leads. Start from the publicly named critics (Turner, Kirsch, and the 20 directors and VPs referenced in Fortune), then expand via co-authorship graphs and public reposts of the CWU and Unite recognition campaign. A sourcing tool that merges GitHub, LinkedIn, and open-web signals will get you further than any single leaked list.

Is this pool realistic for a small startup, or only for Anthropic-scale employers?

Realistic, and arguably easier, for a small values-aligned startup than for a large lab. The bottleneck is Anthropic's hiring capacity, not candidate willingness. If you can put a written defense-avoidance clause in your charter, name your governance model, and grant publication rights, you become the third option in a market where the addressable employer list is very short. The catch is that you have to compete on mission specificity, not comp.

What about US-based DeepMind employees who did not sign?

The US pool of 3,425 DeepMind-linked professionals in Refolk's index is larger than the UK pool, but the organising signal is thinner, so treat it as a longer-cycle inbound market. Non-signers are not disqualified; many agreed privately and chose not to sign for career reasons. First-pass outreach should prioritise the named signers and Turner-proposal co-signers, then work outward to the broader US pool with a slower, referral-driven cadence.

How is this different from sourcing after Project Maven in 2018?

Two things changed. First, the signer count is roughly an order of magnitude larger and includes 20 directors and VPs, which means senior researchers are politically covered in a way they were not in 2018. Second, the UK union drive gives this cohort organisational infrastructure Project Maven signers never had, so the coordinating signals will persist for months rather than weeks. The practical implication is a longer, richer sourcing window and a higher senior-to-junior ratio in the addressable pool.