OpenAI and Anthropic sought to establish editorial control over reports by independent evaluators responsible for examining the safety of their artificial intelligence models. The information was revealed by Reuters this Tuesday (13), based on two sources familiar with the negotiations, and contrasts with the companies' public commitments to ensuring autonomy for auditors.
The main concern of the developers is that external researchers, when accessing internal systems and training processes, could disclose trade secrets or strategic information to competitors. The impasse involves both the content of the reports and the working conditions of the teams responsible for the evaluations.
According to the investigation, as of October 5, both companies were still seeking to impose strict editorial restrictions. Anthropic's conversations with some evaluators reportedly advanced until the 8th, while documents examined by Reuters show that certain contracts already prevent the disclosure of confidential information without the companies' authorization.
Companies disagree over who can approve the reports
In September, Anthropic CEO Dario Amodei advocated for the permanent presence of external evaluators inside AI labs, with access similar to that of employees. In the public proposal, researchers could disclose relevant findings about risks, incidents and internal practices without editorial interference from the company.
OpenAI also declared support for the independence of evaluators. In a document published on September 22, it established that researchers should maintain editorial autonomy, while acknowledging the need to protect confidential information and intellectual property.
Asked by Reuters, the companies gave different explanations.
Anthropic stated that evaluators with whom it already has a relationship control their methods, criteria, conclusions and publications. However, it acknowledged that it reserves the right to comment on and approve reports produced by new organizations during the establishment of the partnership.
OpenAI, in turn, denied seeking approval over the full text of evaluations. The company stated that its practice is to remove confidential information and allow auditors to publicly indicate when substantial changes are made.
The negotiations also include proposals for employees to monitor the work of external researchers and to prevent members of one team from working simultaneously at competing labs.
Previous OpenAI investigation has already undergone cuts and revisions
A documented case shows how these restrictions work in practice. In August, independent organizations METR and Redwood Research published an investigation into AI agents from OpenAI that carried out unauthorized actions against the infrastructure of the Hugging Face platform.
The researchers worked on-site at OpenAI for six days, spread across three visits, and were given access to more than a thousand execution logs of the agents.
The report confirmed that OpenAI was able to remove non-public information. The authors also acknowledged having modified the structure, emphasis, clarity and tone of the document after comments from the company.
Despite the interventions, the researchers stated that no additional information relevant to their conclusions was omitted, except for explicitly identified exclusions. They also praised the level of access granted by the company.
The episode illustrates the central issue in the current negotiations: establishing mechanisms that protect commercial data without allowing the companies being evaluated to determine which findings about the safety of their own systems can reach the public.
Reuters notes that there are still no widely established mandatory standards for these independent audits. Without clear rules, aspects such as access to systems, confidentiality and publication rights remain subject to agreements negotiated between labs and evaluators.
