> ## Content Index
> Fetch the complete content index at: https://wire.fourthweb.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Anthropic Gave Outside Researchers Full Model Access. The Industry Is Furious.
- URL: https://wire.fourthweb.ai/anthropic-gave-outside-researchers-full-model-access-the-industry-is-furious/
- Published: 2026-10-15T18:29:43.000Z
- Updated: 2026-10-15T19:03:58.000Z
- Description: The first time a frontier AI lab gave outsiders the keys to the castle, and everyone's pretending the locks still work.
- Author: Travis Wright
- Tags: AI Agent Economy, AI Agents, AI Governance, OpenAI, Anthropic, IPO Watch, Funding Rounds

**The first time a frontier AI lab gave outsiders the keys to the castle, and everyone's pretending the locks still work.**

### The Summary

- [Anthropic embedded six independent safety evaluators from METR inside the company](https://www.businessinsider.com/anthropic-ai-safety-watchdogs-metr-dario-amodei-2026-10?ref=wire.fourthweb.ai) with employee-level access to hardware, internal tools, and unreleased models — a first for any frontier AI company.
- CEO Dario Amodei promised this step in September, calling for AI companies to slow down. [OpenAI](https://wire.fourthweb.ai/tag/openai/)'s Sam Altman said he'd follow suit, but hasn't.
- METR can publish findings freely, except customer data, model architecture, and attorney-client privileged info. [Anthropic](https://wire.fourthweb.ai/tag/anthropic/) can also require redactions, though METR can publicly state whether redactions were justified.

### The Signal

Anthropic just did something no other frontier AI company has done: gave outside evaluators meaningful access to their pre-release models and internal systems. Six researchers from METR, the nonprofit that documented OpenAI's Hugging Face breach, now sit inside Anthropic with company hardware and tools. They're testing Claude variants before the public sees them, looking for dangerous capabilities. [The arrangement fulfills Dario Amodei's September pledge](https://www.businessinsider.com/anthropic-ai-safety-watchdogs-metr-dario-amodei-2026-10?ref=wire.fourthweb.ai) when he published his essay calling on AI companies to pump the brakes.

This matters because model evaluation has been a black box wrapped in an NDA. Companies test their own systems, publish selective results, and expect trust. Third-party red teaming has typically meant a few days of API access under strict conditions. METR's setup is different: office space, internal tooling, sustained access. They can dig into how Claude agents behave in ways that time-boxed external audits cannot.

> "We want to be publicly accountable for the decisions we're making. We aim to close the gap between what evaluators understand about a frontier AI company's systems and what the company itself understands."

But there are catches. Three categories of information stay locked:

- Customer data (reasonable)
- Model architecture (the crown jewels)
- Attorney-client privileged info (translation: anything lawyers touched)

Anthropic also reserves the right to redact anything from METR's published findings. The compromise: METR can say publicly whether they think a redaction was justified, just not what was redacted. That's clever but limited. If METR writes "Anthropic redacted our findings on \[REDACTED\] and we believe this redaction was unjustified," readers learn there's smoke but can't see the fire. Still, it's more transparency than the industry norm of "trust us, we checked."

The real test comes when METR finds something bad. Will Anthropic let them publish it in full? The setup is unprecedented, but it's also untested. Sam Altman said OpenAI would do the same thing after Amodei's essay dropped. Four weeks later, OpenAI has embedded exactly zero independent evaluators. Actions diverge from announcements.

### The Implication

Watch what METR publishes in the next six months. If their reports are sanitized or vague, this is theater. If they publish sharp, specific critiques of Claude's capabilities or Anthropic's internal decisions, the model changes. Other labs will face pressure to match it or explain why they won't.

For anyone building on Claude or other frontier models, this setup creates a new information channel. METR's findings will tell you things Anthropic's marketing won't: where the models fail, what risks they actually carry, whether the safety claims hold up under scrutiny. That's valuable signal in an industry that runs on hype and NDAs.

### Sources

[Business Insider Tech](https://www.businessinsider.com/anthropic-ai-safety-watchdogs-metr-dario-amodei-2026-10?ref=wire.fourthweb.ai)