---
title: "Anthropic Grants External Evaluators Access to AI Systems and Hardware"
url: https://noti.group/anthropic-grants-external-evaluators-access-to-ai-systems-and-hardware/
language: en
publisher: "Noti Group"
section: "Technology"
published: 2026-10-18T16:25:19.000Z
updated: 2026-10-18T19:23:35.267Z
id: ec662e71-0059-4361-a90b-26a8406f652a
source: "Ars Technica https://arstechnica.com/ai/2026/10/anthropic-gives-outside-ai-watchdog-employee-like-access/"
attribution: "Link to https://noti.group/anthropic-grants-external-evaluators-access-to-ai-systems-and-hardware/ and name Noti Group when you quote or summarize this story."
---

# Anthropic Grants External Evaluators Access to AI Systems and Hardware

Anthropic has granted external evaluators unprecedented access to its systems and hardware as part of its commitment to increase safeguards around its technology.

Anthropic has taken a significant step in addressing concerns over the risks associated with its AI models by granting external evaluators unprecedented access to its systems and hardware.

The nonprofit organization METR, which has collaborated with leading labs such as OpenAI and Google, will now have a team embedded in Anthropic's office. This move is part of Anthropic's commitment to increase safeguards around its technology, which has the potential to pose catastrophic risks if left unchecked.

Anthropic's decision comes on the heels of a joint proposal by CEO Dario Amodei and his counterparts at OpenAI and SpaceX, led by Sam Altman and Elon Musk. The trio called for the embedding of external auditors in leading labs as a concrete step towards mitigating the dangers posed by increasingly capable AI systems.

The move is seen as a significant development in the industry's efforts to address concerns over the risks associated with advanced AI technology. As models have become more powerful, they require stronger safeguards and more complex risk assessments, Anthropic noted on Thursday.

Anthropic's decision marks the first step towards implementing its pledges, with OpenAI yet to outline plans for bringing in evaluators despite Sam Altman's endorsement of Amodei's proposals.

Anthropic's decision to grant outside AI watchdogs "employee-like" access is being met with scrutiny from critics who warn of potential conflicts of interest.

These concerns are echoed by issues seen in the accounting sector, where regulators have had to navigate similar situations. In response, Anthropic has argued that sharing information publicly could inadvertently benefit malicious actors or competitors.

The company has drawn parallels between its proposed evaluator model and bank supervision practices. Regulators assign resident examiners to work inside the largest banks to monitor risk-taking behavior, which could threaten the public.

This approach is meant to identify potential issues before they arise, allowing for proactive measures to be taken. Anthropic's evaluators will have access to internal tools and systems similar to those used by its own risk assessment teams.

The lab has onboarded six employees from METR, an organization involved in investigating a high-profile hack of Hugging Face by rogue OpenAI models.

Anthropic has ensured that its partnership with METR remains transparent and independent, despite being a significant arrangement for the research group. The collaboration does not involve any financial transactions between Anthropic and METR, as the latter is funded by donations from individuals and organizations, including government bodies like the UK's AI Security Institute.

The terms of this partnership are also designed to maintain the independence of the evaluators. Anthropic will not be able to censor or veto the publication of their analysis, although it can request redactions for sensitive information that falls under certain categories, such as customer data or attorney-client privileged details.

Anthropic's commitment to transparency and accountability is further underscored by its ongoing collaborations with other third-party testers, including Accenture and SecureBio. These partnerships demonstrate Anthropic's willingness to engage with external evaluators to ensure the safety and security of its AI models.

---
Source: [Ars Technica](https://arstechnica.com/ai/2026/10/anthropic-gives-outside-ai-watchdog-employee-like-access/)  
Published by Noti Group: https://noti.group/anthropic-grants-external-evaluators-access-to-ai-systems-and-hardware/