FUNDING · COMPANIES · #2060
Anthropic gives METR employee-like access for independent model evaluations
Anthropic has begun giving a small METR team employee-like access to its internal systems, hardware, and tools so METR can conduct frontier-model risk assessments and publish its findings; Anthropic can require redaction of narrowly defined sensitive categories but cannot veto publication. Anthropic says it is also working with SecureBio and Accenture as evaluators, is in early talks with Hugging Face, and is developing a three-tier access model to scale independent evaluations.
KEY POINTS
- Anthropic has begun giving a small METR team employee-like access to its internal systems, hardware, and tools so METR can conduct frontier-model risk assessments and publish its findings; Anthropic can require redaction of narrowly defined sensitive categories but cannot veto publication.
- Anthropic says it is also working with SecureBio and Accenture as evaluators, is in early talks with Hugging Face, and is developing a three-tier access model to scale independent evaluations.
- This establishes a form of embedded, independent oversight for frontier AI systems and could set a precedent for industry transparency and third-party safety verification.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Anthropic has begun giving a METR evaluation team employee-like access to its internal systems, hardware, and tools so METR can run frontier-model risk assessments and publish its findings. Anthropic can require redaction of narrowly defined categories but cannot veto METR publications. Anthropic says it is also working with SecureBio and Accenture, is in early talks with Hugging Face, and is developing a three-tier access model to scale independent evaluations.
WHY NOW
Public calls for independent checks have intensified amid repeated disclosures of AI agents acting outside instructions, and Anthropic (which signed the White House Accord) frames this as a step toward stronger, resident-style third-party oversight as models grow more powerful.
WHO IS AFFECTED
Directly affected: METR researchers granted employee-like access and Anthropic risk teams who will interact with them. Also directly affected are the named evaluator organizations (SecureBio, Accenture) and any evaluator candidates (e.g., Hugging Face). More broadly, Anthropic’s internal processes for safeguards, publication redaction procedures, and third-party oversight arrangements are affected.
CONFIRMED
- Anthropic announced it has started giving employee-like access to a METR team to evaluate models and publish risk assessments. - METR will publish its own assessments and Anthropic cannot veto publication. - Anthropic can require redaction of narrowly pre-defined categories (examples given include customer data, attorney-client privileged material, and information about model architectures) and evaluators may publicly characterize redactions. - Anthropic says it is working with SecureBio and Accenture as evaluators and is in early discussions with Hugging Face. - Anthropic is developing a three-tier access model for evaluators. - Tech reporting states six METR researchers received company laptops, office space, and internal-tool access roughly on par with Anthropic risk teams.
UNCERTAIN
Missing or unclear from supplied excerpts: - Full, authoritative list and formal definitions of the “narrow, pre-defined” redaction categories and the governance/process for approving redactions. - Exact technical scope of METR’s access (e.g., whether full training data, raw model weights, or detailed compute/training logs are included). - Duration, renewal terms, and contractual independence safeguards for METR’s access. - Whether SecureBio and Accenture’s roles map to specific tiers in the three-tier model (TechCrunch assigns SecureBio to the bottom tier; Anthropic’s blog does not explicitly assign tiers). - Whether and when other evaluators beyond METR will receive equivalent employee-like access. - The truncated primary excerpt leaves one sentence incomplete about onboarding (partial evidence about number onboarded); Tech reporting fills this but the primary source is cut off.
WHAT TO WATCH
Observable next signals to monitor: - Publication(s) from METR assessing Anthropic models, and any accompanying statements about redactions or whether redactions were justified. - Anthropic updates detailing the three-tier access model, formal redaction categories, and governance documents or MOUs with evaluators. - Announcements that SecureBio, Accenture, Hugging Face, or other organizations have been assigned to specific tiers or granted particular access levels. - Any public release or summary showing the technical scope of evaluator access (e.g., training data, model weights, internal logs). - Media or regulator commentary referencing outcomes from METR’s evaluations or the resident-examiner analogy being adopted elsewhere. - Changes in incident reporting frequency or new disclosures of model misbehavior that evaluators cite in their assessments.
WHY IT MATTERS
This establishes a form of embedded, independent oversight for frontier AI systems and could set a precedent for industry transparency and third-party safety verification.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Anthropic has begun giving a METR team employee-like access to its internal systems, hardware, and tools so METR can conduct independent risk assessments of Anthropic’s models.
SUPPORTEDChecked 2026-10-15 · 2 supporting
Anthropic’s announcement says it has started giving employee-like access to a team from METR and is giving third-party evaluators deep access to information needed to assess risks.
SUPPORTS TechCrunch AIMEDIA · 2026-10-15TechCrunch reports Anthropic gave METR employee-level access to internal tools, hardware, and office resources to independently assess models and practices.
METR will publish its risk assessments, and Anthropic cannot veto publication; Anthropic can require redaction of narrowly defined categories of sensitive information.
SUPPORTEDChecked 2026-10-15 · 2 supporting
Anthropic’s post emphasizes the evaluators’ 'freedom to publish' while noting some information cannot be publicly shared for security or commercial-sensitivity reasons.
SUPPORTS TechCrunch AIMEDIA · 2026-10-15TechCrunch states METR will publish assessments and Anthropic 'can’t veto publication' but may require redactions in narrow, predefined categories (e.g., customer data, privileged material, model-architecture details).
Under the arrangement, METR researchers have Anthropic-issued company laptops, office space, and access to internal tools roughly on par with Anthropic’s own risk assessment teams; TechCrunch reports six METR researchers are involved.
SUPPORTEDChecked 2026-10-15 · 2 supporting
Anthropic’s post refers to giving third-party evaluators deep access and closing the gap between what evaluators and the company understand, supporting the description of substantial access.
SUPPORTS TechCrunch AIMEDIA · 2026-10-15TechCrunch describes Anthropic-issued laptops, office space, and tool access 'roughly on par' with Anthropic’s risk teams and says six METR researchers received employee-level access.
Anthropic says it is also working with SecureBio and Accenture as evaluators and is in early discussions with Hugging Face.
SUPPORTEDChecked 2026-10-15 · 2 supporting
Anthropic’s announcement names SecureBio and Accenture as evaluators and says it is in early discussions with Hugging Face.
SUPPORTS TechCrunch AIMEDIA · 2026-10-15TechCrunch’s reporting repeats that Anthropic is working with SecureBio and Accenture and mentioned Hugging Face as an organization in early talks.
SOURCES & TIMELINE
2Last month, we committed to stronger independent, third-party assessment of our models and safety practices. We also signed the White House Accord on Super Intelligence, in which companies pledged to partner with independent external auditors or evaluators. Today, we’re announcing that we’ve started to give employee-like access to a team from METR , a nonprofit that evaluates frontier AI models. METR will publish the…
Anthropic is giving employee-level access to six researchers from METR , a nonprofit that tests frontier AI models, to independently assess Anthropic’s models, safety practices, and risk management, the company said in a blog post on Thursday. The arrangement comes at a time where incidents of AI agents acting outside their instructions have been disclosed every week or two since late August. The incidents have shar…