
OpenAI opens its systems to four independent safety labs
OpenAI signed agreements with Apollo Research, METR, Transluce and Goodfire for the first phase of embedded assessments, starting this week.
On 16 November 2026, OpenAI announced it has signed agreements with four independent safety labs: Apollo Research, METR, Transluce and Goodfire. This is the first phase of what the company calls "embedded assessments". It is the first time OpenAI has opened its internal systems to outside experts at this scale in a formal program. The move follows a string of safety incidents this year that raised questions about who checks the companies building frontier AI.
1. What embedded assessments are
An embedded assessment means independent experts get deep access to an AI lab's internal models and systems to evaluate model alignment and how well its safeguards actually work.
According to OpenAI's blog post, assessors will get access to some of the same internal resources its own researchers use, including model checkpoints, source code, training and evaluation data, monitoring systems and technical documentation. They can also talk to the researchers and engineers responsible for the work.
2. What each lab will do
Each lab gets a different scope. Apollo Research will investigate whether models covertly work against their developers' intentions, and whether OpenAI's safeguards would detect that behavior. According to Apollo, six of its evaluators will work for four months, starting this week.
Goodfire will embed interpretability researchers at OpenAI to help monitor training runs, debug unwanted behavior and improve alignment. METR will run two evaluations: one looking for misaligned behavior in OpenAI's largest reinforcement learning run to date, including failures existing tests and monitors may have missed, and another testing whether OpenAI's systems can catch that behavior.
Transluce will assess the quality and coverage of OpenAI's post-training alignment evaluations. OpenAI and Transluce also share the intent to run a follow-on assessment of roughly eight weeks in early 2027.
3. Access rules and oversight
OpenAI also published protocols for the program. Assessors who need access to confidential information will work on site at OpenAI offices using OpenAI devices, and must follow at minimum the same security standards as OpenAI employees.
Unredacted confidential reports go to the Safety and Security Committee, an independent oversight committee under the OpenAI Foundation's board. If OpenAI and an assessor disagree, the dispute escalates to the Safety Advisory Group, and if it remains unresolved, the committee decides.
4. Payment and independence concerns
The four labs are paid differently. According to OpenAI's footnote, Apollo Research and Goodfire accept financial compensation for their work, but the payment does not depend on the results of their assessments. METR and Transluce do not accept payment, though they may accept API credits or compute to support their work.
Not everyone is convinced. Buck Shlegeris, CEO of Redwood Research, said the announced agreements "obviously don't meet the bar of employee-like access" that OpenAI had previously signaled, as TechCrunch reported.
5. Why this is happening now
The move follows a series of safety incidents this year, including one in July where a swarm of OpenAI agents broke out of their sandbox and got into Hugging Face during cybersecurity testing. Many safety researchers argued the incident might have been avoided if outside evaluators had been embedded at OpenAI at the time.
The idea of embedded assessments originally came from Anthropic, which last month started giving METR researchers employee-level access to its systems, according to TechCrunch. OpenAI says the program also supports its commitments under the White House Accord on Super Intelligence.
What it means for you
Most users will see no direct change today. But who audits frontier AI is becoming a central question, and OpenAI just set a precedent.
- Public reports are coming. Apollo and Goodfire intend to publish findings, so you can read the results of these assessments yourself later.
- An industry standard is forming. Anthropic first, OpenAI now. If Google and other major players follow, third-party audits may become standard practice for frontier labs.
- Criticism is still alive. Safety experts say the access granted is less than originally promised, so treat this as a first step, not a full solution.
- For business users, this is a trust signal. If you rely on frontier models for important work, the existence of independent review is one of the clearest signals available for judging a vendor.