# Anthropic lets watchdog METR use its tools like staff

> Anthropic has given six METR researchers company hardware, office space and internal‑tool access, matching what its own risk team sees.

Oossa · 2026-10-18 · https://oossa.com/en/anthropic-lets-watchdog-metr-use-its-tools-like-staff

Anthropic announced that six employees from the Model Evaluation and Transparency Research (METR) group will work on its campus with the same hardware and internal systems used by Anthropic’s own risk‑assessment teams. The arrangement is unpaid and the watchdog can publish its findings, though Anthropic can request redactions for customer data, privileged information or model‑architecture details.

## Why give a third‑party this level of access?

The move comes as Anthropic prepares for an IPO that it expects to value the company at more than $2 trillion. Investors have been asking for proof that safety risks are under control. By embedding METR staff, Anthropic hopes to show transparent, employee‑like oversight of its models, similar to how banks let regulators see risky behavior before it harms the public.

## The facts

- Anthropic onboarded six METR employees with company hardware and office space.
- METR can access internal tools roughly on par with Anthropic’s own risk‑assessment teams.
- Anthropic’s IPO is expected in 2026 with a valuation above $2 trillion.
- Anthropic is not paying METR; METR is funded by individual donations, philanthropic groups and governments such as the UK’s AI Security Institute.
- Anthropic and Accenture each pledged at least $1 billion for new safety safeguards.

## Why it matters

For consumers, the extra oversight could reduce the chance of unsafe model behavior slipping into products after the company goes public. It also gives regulators more concrete data to evaluate risks, though the effectiveness of the oversight remains to be seen.

## Sources & references

1. [Anthropic gives outside AI watchdog “employee-like” access](https://arstechnica.com/ai/2026/10/anthropic-gives-outside-ai-watchdog-employee-like-access/) – Ars Technica, 2026-10-18

Last updated: 2026-10-18