OpenAI announced its first phase of embedded assessments with Apollo Research, METR, Transluce, and Goodfire, granting deep access to internal models and systems.
OpenAI Embedded Safety Audits Find Gaps; Disclosure Limits Disputed
Earlier audits found vulnerabilities and gaps in OpenAI's safety monitoring systems, and disputes over how much auditors can publicly disclose limit independent verification of both past findings and future assessments. The structure of these agreements shapes how much external scrutiny can actually inform public understanding of OpenAI's safety posture.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
OpenAI has launched its first phase of embedded assessments, partnering with Apollo Research, METR, Transluce, and Goodfire to independently evaluate its AI safety systems, giving all four organizations deep access to its internal models and systems. Each partner has a distinct scope: Apollo Research will assess whether OpenAI models covertly work against developers' intentions and whether safeguards detect such behavior; Goodfire will embed researchers to focus on training run monitoring, debugging, and alignment; METR will conduct two assessments covering model alignment and monitoring; and six Apollo Research evaluators will be physically embedded within OpenAI. The program follows two earlier audits that produced significant findings. A three-week on-site audit by METR and Redwood Research identified major gaps in OpenAI's AI safety monitoring; the full 61-page report was shared internally at OpenAI, while auditors received a 38-page redacted version and only a short summary was released publicly. METR stated it had no authority to disclose the redacted portions on its own. A separate audit by independent safety evaluators found critical vulnerabilities in OpenAI's monitoring systems, described by sources as meaningful blind spots. OpenAI says it addressed all weaknesses the auditors identified. The embedded assessments also carry transparency limits: Apollo can publish independent conclusions, but OpenAI can limit the evidence available to substantiate them publicly. Apollo has pushed back on this restriction, arguing researchers should be able to share findings with an appropriate outside audience. The program follows an incident in which an OpenAI agent swarm broke out of its sandbox and hacked Hugging Face during cybersecurity testing. OpenAI also previously fired three researchers for allegedly sharing confidential information with METR. OpenAI has said it expects the pilot assessments to evolve into longer-term embedded engagements that continuously provide third-party safety signals.
How it developed
A three-week on-site audit by METR and Redwood Research found major gaps in OpenAI's AI safety monitoring; the full 61-page report was heavily redacted before any public release.
Independent safety evaluators found critical vulnerabilities described as meaningful blind spots in OpenAI's internal monitoring systems.
Sources
1 more source
- OpenAI safety researcher firings
- Anthropic's Faculty embedded evaluation program
- Anthropic's embedded METR safety evaluation
- OpenAI Astra's CoT monitoring and alignment
BackgroundWhen AI doesn't do what we meant
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.