Anthropic Gave a Safety Watchdog Staff-Level Access. A Viral Thread Says That Watchdog's Funders Hold Anthropic Stock
This follows a September pledge from CEO Dario Amodei to embed independent evaluators inside Anthropic with real access, part of Anthropic's commitments under the White House Accord on Superintelligence, where AI companies agreed to work with outside auditors. This update is the first concrete sign of what that pledge actually looks like in practice, rather than a promise.
The access tiers Anthropic actually built
| Tier | Scope | Publication rights |
|---|---|---|
| 1 | Narrowly scoped investigations, access chosen by Anthropic | Requires Anthropic's approval before publishing |
| 2 | Broader information access | Can publish without approval; Anthropic can still require redaction of set categories |
| 3 (METR, currently) | Employee-like access: hardware, office space, internal tools and systems | Can publish without approval, similar redaction limits apply |
METR's six onboarded staff can't see everything. Anthropic carves out customer data, attorney-client privileged material, and model architecture details as off-limits, and it can require redaction of defined categories like confidential IP before anything is published, though METR can state publicly whether it thinks a requested redaction was justified. Anthropic also says it has begun similar, narrower arrangements with SecureBio on biological and chemical weapons risk review, and has planned partnerships with Accenture and Faculty on red-teaming and alignment assessment, with early talks underway with other groups including Hugging Face.
Anthropic's own framing: 'We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same,' per its announcement.
The independence fight this arrangement walked into
In the weeks before this update, a widely circulated thread from researcher Kevin Bass argued that METR operates inside what he calls an 'Anthropic Network': a set of organizations connected by shared funding, where Good Ventures (funded by Dustin Moskovitz) holds Anthropic stock and channels money through Open Philanthropy into parts of the AI-safety ecosystem that includes METR's own funders, per coverage of the thread. Bass has called for a US congressional investigation into the ties.
METR's response, as reported, is that it does not accept funding directly from AI companies or their employees, and that its funders have no say over which projects it takes on or what it publishes. One independent analysis on LessWrong accepts that defense but adds a narrower point worth sitting with: METR does accept free compute credits from the labs it evaluates, including Anthropic, which isn't a cash payment but is still a resource dependency, even if the same analysis calls METR's position on it ethically reasonable, per the LessWrong write-up.
Why the access grant is a real step
What this update actually adds
- ✦Staff-level access is more than a review-the-paperwork arrangement
- ✦Publication rights without pre-approval let METR say things Anthropic might not want said
- ✦First test of a tiered framework other labs could be pressured to match
Why skepticism is still fair
What the independence critique gets right
- ✦Compute credits from Anthropic to METR are a real dependency, even if not cash
- ✦Funding-network overlaps are documented, not fabricated, even if they don't prove direction
- ✦Anthropic still defines the tiers and the redaction categories METR operates inside
What would actually settle it
The missing evidence on both sides
- ✦A case where METR publishes something Anthropic clearly didn't want public
- ✦Independent financial disclosure from METR beyond its own characterization
- ✦A second lab matching Anthropic's Tier 3 arrangement, showing it isn't just one company's PR move
Neither side's claim is fully settled by what's public right now. The funding overlaps Bass documented are real and traceable; so is METR's stated policy against taking AI-company money directly. What nobody has shown yet is a concrete instance of Anthropic's influence, direct or indirect, actually shaping what METR did or didn't say about Anthropic's own models.
What access did Anthropic actually give METR?
'Employee-like access' to Anthropic's hardware, office space, and internal tools and systems, roughly matching what Anthropic's own internal risk-assessment teams get, with the right to publish findings without Anthropic's pre-approval, subject to narrow redaction categories.
Is METR actually independent of Anthropic?
It's disputed. METR says it takes no direct funding from AI companies or their employees and that funders don't direct its projects. Critics point to documented financial ties running through Good Ventures and Open Philanthropy into the broader AI-safety funding ecosystem, plus the free compute credits METR accepts from the labs it evaluates, including Anthropic.
What can't METR see or publish?
Customer data, attorney-client privileged material, and model architecture details are off-limits. Anthropic can also require redaction of defined categories like confidential IP before publication, though METR can state publicly if it disagrees with a redaction request.
Why is Anthropic doing this now?
It follows Dario Amodei's September pledge to embed independent evaluators with real access, part of Anthropic's commitments under the White House Accord on Superintelligence, where AI companies agreed to work with outside auditors.
Are other AI companies doing anything similar?
Anthropic says it's in early discussions with other evaluators, including Hugging Face, and has separate narrower arrangements planned with SecureBio, Accenture, and Faculty. No other major AI lab has publicly matched Anthropic's Tier 3, staff-level access arrangement yet.
Try the tools ✦
Free browser tools that never upload your files.