Anthropic CEO Dario Amodei published an essay on September 12. He said Anthropic would give independent evaluators permanent access to verify safety practices. The evaluators could publish key findings about risk levels, incidents and practices. They could do so without editorial control by Anthropic. OpenAI CEO Sam Altman endorsed the idea days later. He said assessors should have editorial independence.
Reuters reported on October 13 that as recently as October 5, both companies were seeking strict editorial control. They wanted approval rights over what the evaluators could publish. Two people familiar with the discussions said Anthropic and OpenAI wanted approval rights over reports. By October 8 the talks had evolved. Anthropic told Reuters that for new evaluator firms, it reserved the right to comment on reports. It also reserved the right to approve them.
The API grants the access it does not document the control
Atelier routes prompts to seven model providers. The estimating tool calls an API to score bids. The flight planner queries weather and fuel data. Each integration grants an API key. The key allows the platform to send a request. The provider returns a response. The API documentation describes rate limits, input formats and response schemas. It does not describe what the provider logs. It does not say how long it keeps the logs or who can access them. The API does not say whether requests train models.
The September essay said evaluators would get employee-level access. They would receive desks, badges, laptops and permissions comparable to internal risk teams. They could publish findings about risks, incidents and the access they received. Amodei said Anthropic would reserve a narrow ability to redact security-sensitive or confidential information. The October negotiations included embedding lab staff to observe evaluators while they worked. They included clauses preventing evaluators from working at competing labs. The contracts gave the company approval rights over reports from new evaluator firms.
The grant and the control are separate fields
The construction ERP processes invoices. The permit app parses zoning codes. The dating-profile tool runs on one Cloudflare Worker. Each platform calls a third-party API. The API accepts the request if the key is valid. It returns the response. The integration assumes the response is clean. The API terms of service sit in a separate document. The terms describe data retention, sharing policies and future use. The call site shows no oath.
Anthropic named Faculty as its first embedded evaluator in September. Faculty is described as Accenture's AI division. The evaluators would inspect training, deployment and safety commitments. They would test adherence to responsible scaling policies. The company said evaluators it had previously partnered with controlled their methods, criteria, conclusions and publications. For new firms, it reserved approval rights. It said the firms needed time to learn about its models and establish a working relationship. The distinction between existing and new evaluators appeared in the October statement. It did not appear in the September essay.
The key worked. Approval was separate.
The jewelry site processes drawing orders. Curbscript generates service routes. The dog-training site runs review gates. Each integration reads the API reference. The reference shows the endpoint, the required headers and the expected response. It does not show the clauses that govern what the provider can do with the request. Those clauses live in the terms of service or a data-processing agreement. The developer who writes the integration may not read those documents. The API grants access. The contract reserves control. The access and the control are separate fields in separate documents. The September essay described the access. The October contracts described the control.