Anthropic AI evaluators now include an embedded six-person team from nonprofit research group METR, giving outsiders employee-like access to examine frontier-model risks and safety practices.

Anthropic AI Evaluators Gain Deeper Access

Anthropic said on October 15 that it had begun providing employee-like access to six METR staff members. The evaluators received company hardware, workspace and access comparable to Anthropic’s internal risk-assessment teams, according to the company’s official announcement.

Anthropic excluded customer information, legally privileged material and most model-architecture details from routine access. Employees can otherwise share relevant safety information with the embedded team as they would with internal alignment-risk staff.

The structure responds to a persistent problem in AI oversight: developers know more about their systems than external reviewers. Recent corporate research has also shown gaps between incident escalation and policy changes, including IBM’s EMEA AI governance findings.

Publication Rights Limit Company Control

METR will evaluate Anthropic’s risk levels, incidents and practices. Anthropic said it cannot veto publication, although it can require redactions for narrow, predefined categories of confidential information.

METR can publicly describe whether it believes a redaction was justified without disclosing the protected material. That condition is designed to preserve evaluator independence while reducing the risk of exposing customer data, trade secrets or information useful to attackers.

The arrangement matters as developers release more capable models for autonomous work. Nvidia’s recent Nemotron 3.5 agent models show how quickly systems are moving onto servers, local workstations and edge devices.

A Three-Tier Oversight Model

Anthropic outlined three access levels. First-tier evaluators receive information for narrowly scoped investigations and need company approval to publish. Second-tier evaluators gain more information and can publish without approval, while third-tier teams receive the embedded access granted to METR.

SecureBio currently occupies the first tier and is reviewing the chemical and biological weapons section of Anthropic’s August Risk Report. Anthropic is also working with Accenture and Faculty on model evaluations, red teaming, alignment assessments and safeguard testing.

The program does not eliminate conflicts between transparency and security. It creates a framework for qualified outsiders to test claims with more evidence than public users receive, while maintaining controls around sensitive systems.

External evaluation is one layer of risk management. Enterprises must also secure the infrastructure used by AI agents, including protections such as Radware’s MCP server controls. The effectiveness of Anthropic’s model will depend on what evaluators can inspect and whether published findings influence product decisions.

Sharron Kendi is a seasoned crypto market analyst and writer with over three years of experience decoding blockchain trends, price movements, and market dynamics. She holds a Bachelor’s Degree in Commerce (Finance) from Kenyatta University, blending a solid academic foundation with a sharp eye for technical analysis and a deep understanding of on-chain data. Her work delivers clear, data-driven insights that empower investors to navigate the fast-evolving digital asset space with confidence.