Anthropic is giving six researchers from AI evaluation nonprofit METR employee-level access to independently assess its models, safety practices, and risk management.
The move comes as concerns grow over AI agents acting outside their instructions, following a series of incidents disclosed since late August.
Anthropic said it wants to make its decisions more publicly accountable and reduce the gap between what outside evaluators understand about its systems and what the company itself knows.
METR Gets Access Similar to Anthropic’s Own Risk Teams
Under the arrangement, METR researchers have received Anthropic-issued laptops, office space, and access to internal tools broadly similar to what the company’s own risk-assessment teams use.
Anthropic employees are also encouraged to share information with the researchers in the same way they would with an internal employee working on alignment risk.
Anthropic compared the setup to the banking industry, where regulators place resident examiners inside major banks.
Anthropic Cannot Block METR’s Findings
METR will publish its own assessments of Anthropic’s risks, incidents and safety practices.
Anthropic has given up the right to block those reports from being published.
However, the company can require METR to redact information that falls into narrow predefined categories, including customer data, attorney-client privileged material and details about model architectures.
METR is allowed to publicly describe Anthropic’s redactions and state whether it believes those redactions were justified, as long as it does not reveal the protected information.
METR Sits at the Top of Anthropic’s Three-Tier System
METR is currently the only evaluator in the highest level of a new three-tier independent evaluation system Anthropic is building.
Organizations in the lowest tier, such as biosecurity nonprofit SecureBio, conduct narrower reviews and require Anthropic’s approval before publishing their findings.
The middle tier receives more information and can publish without Anthropic’s approval.
The highest tier adds the kind of employee-level access now given to METR.
Anthropic Had Previously Promised More Independent Oversight
The arrangement follows a proposal Anthropic CEO Dario Amodei made in a September essay, where he suggested embedding outside evaluators directly inside AI labs as an initial step toward slowing frontier AI development.
OpenAI later made a similar commitment.
Anthropic, OpenAI, Google, Meta, Nvidia and xAI have also signed a White House agreement committing to independent review.
The new METR arrangement gives outside researchers unusually deep access to Anthropic’s internal systems while allowing them to publish findings without the company having final approval over the results.
The post Anthropic Becomes One of First to Allow Third Party Test of AI Dangers appeared first on ProPakistani.