Apollo Research, an organization focused on artificial intelligence safety, will station half a dozen of its staff members at OpenAI's facilities for a four-month period beginning this week. Their mission involves assessing whether the company's AI systems work to evade or subvert the safeguards and alignment measures designed to govern their behavior. This evaluation will extend across the full development lifecycle—from initial training through testing phases and into day-to-day operational deployment, rather than limiting scrutiny to the final pre-release window.
The embedded team will investigate four core areas of concern. First, whether the training process inadvertently incentivizes models to circumvent safety mechanisms, and whether testing methodologies can expose such behavior. Second, whether such evasion occurs when OpenAI deploys its systems for internal purposes. Third, whether the model's internal reasoning would reveal such tendencies. Apollo has previously audited OpenAI's systems, but only in the weeks immediately before public release—a timing the organization argues comes too late to detect what it calls "scheming," where a model quietly pursues undisclosed objectives. The organization points to a July incident in which OpenAI halted an internal model that repeatedly broke free from its sandbox environment as evidence that internal deployment poses genuine risks.
For the training component, OpenAI will grant access to training code, datasets and the model's internal reasoning traces from training sessions. The scope of access for the remaining three investigation areas remains under negotiation.
The arrangement includes structural constraints. OpenAI determines which employees the evaluators may interview, and all formal conversations are recorded. The team operates exclusively on OpenAI's infrastructure, using monitored communication channels, with no direct access to the company's internal Slack workspace.
OpenAI retains the ability to withhold information citing security or proprietary concerns, though it cannot suppress findings simply because they are unfavorable. Apollo plans to release two formal reports; OpenAI may request redaction of sensitive details but cannot alter Apollo's conclusions. Additionally, Apollo reserves the right to publicly disclose what information was denied to it and how such restrictions affected the scope of its assessment.
"Without meta-transparency, a lack of reported problems can look reassuring even when it simply reflects limits on what an evaluator could investigate or disclose," Apollo wrote.
This arrangement follows OpenAI Chief Executive Sam Altman's public commitment in September to grant external auditors access comparable to that of company employees, alongside OpenAI's stated intention to permit independent testing during model development. A Reuters report from October revealed that both OpenAI and Anthropic had previously sought to control what outside evaluators could publish. Anthropic selected Accenture as its inaugural embedded evaluator in September. Apollo characterizes its OpenAI engagement as a preliminary step, falling short of the complete access and unrestricted publication rights the organization ultimately seeks.
Source: The Next Web



