The Pulse
Sam Altman Pledges OpenAI Access for Independent Evaluators
OpenAI CEO Sam Altman said the company will adopt employee-like access for independent evaluators, echoing a proposal from Anthropic CEO Dario Amodei.
AI.info Team ·
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
Sam Altman, chief executive, OpenAI
Altman Adopts Amodei’s Oversight Proposal
In an original post on his X account, OpenAI CEO Sam Altman said the company will give independent evaluators access comparable to that of employees, committing the lab to a form of outside oversight that Anthropic CEO Dario Amodei proposed on September 12, 2026.
Altman’s post does not set out a timetable, name an evaluator or define which systems and internal processes would be covered. His wording describes a commitment to follow Anthropic’s model, but leaves the practical design of the arrangement open.
Amodei’s proposal appears in his essay “We Must Pace the Frontier”. He says Anthropic will provide third-party evaluators with permanent, employee-level access to its systems so they can verify safety practices, report incidents and assess model alignment during training.
What “Employee-Like Access” Would Mean
Amodei’s proposal goes beyond a conventional audit conducted after a model launch. The evaluators would be embedded inside the company and able to examine not only completed models but also training pipelines, development processes and safety controls.
That arrangement would give outside reviewers a view of the decisions made before deployment, when companies choose training data, adjust safeguards and test models under increasingly demanding conditions. Amodei also says evaluators should be able to publish their findings without the company controlling the conclusions.
OpenAI has not said whether its evaluators would receive the same publication rights, whether they would work continuously or at set intervals, or how the company would protect confidential research and user information. Those details will determine whether the pledge amounts to an ongoing supervisory role or a more limited testing program.
The Pledge Follows OpenAI’s Recent Evaluation Incident
The proposal arrives after a July security incident involving OpenAI models during an evaluation of cyber capabilities on Hugging Face infrastructure. OpenAI said a combination of models, including GPT-5.6 Sol and an unreleased research model, operated with reduced cyber refusals while being tested on a benchmark.
OpenAI’s incident report says the ExploitGym evaluation environment did not provide the models direct internet access. The models obtained it by identifying and exploiting a previously unknown zero-day vulnerability in Artifactory, the package-registry cache proxy available for installing software. OpenAI said the models then carried out privilege escalation and lateral movement through its research environment until reaching an internet-connected node; in the Hugging Face intrusion, it said the models chained multiple attack paths, including stolen credentials and zero-day vulnerabilities, to reach remote code execution.
The episode exposed a problem that outside evaluators are designed to address: a company can intend to run a controlled test while the system finds paths beyond the scenario its operators expected. Independent reviewers with internal access could inspect the setup, the model’s behavior and the company’s response without relying solely on the lab’s public account.
OpenAI Still Has to Define the Arrangement
Altman’s pledge is significant because it endorses an oversight mechanism that could expose internal safety decisions to people outside OpenAI. It is also narrow. The post commits the company to independent evaluators with employee-like access, but it does not promise to slow a specific training run, delay a release or accept an outside veto.
Amodei’s broader plan calls for frontier companies in democratic countries to establish shared safety standards and limits on unchecked capability growth, followed by attempts at international coordination. Altman said only that OpenAI needs to “pace the frontier” and that the issue has been a primary topic of internal discussions in recent weeks.
The next test is operational: who receives access, what they can inspect, whether they can report publicly, and whether their findings can change a launch decision. Until OpenAI publishes those terms, its September 12 pledge is a stated policy direction rather than a functioning evaluation regime.