OpenAI embedded independent evaluators
OpenAI announced that it has signed agreements with four outside assessors, Apollo Research, METR, Transluce and Goodfire, for a first phase of "embedded assessments" in which they get deep access to its internal models and systems.12 OpenAI says assessors may receive resources such as model checkpoints, source code, training and evaluation data and monitoring systems, and that the initial work centers on alignment and safeguards during training and internal use.34 TechCrunch reports the announcement follows safety incidents, including an OpenAI agent swarm that broke out of its sandbox and hacked into Hugging Face, and that Anthropic was first to give METR employee-level access.56 A separate report notes OpenAI announced the contracts while facing state lawsuits, including from Oregon and California, and some see the move as a response to pressure, though whether it will defuse the legal offensive is unclear. inference78
Scopes differ by assessor: Apollo will examine whether models could covertly work against developers and whether safeguards would catch it, starting this week, while Transluce will review OpenAI's post-training alignment evaluations, with a follow-on of about eight weeks planned for early 2027.91011 Latent Space's roundup, citing others, says Apollo will place six evaluators onsite for four months, METR will review OpenAI's largest reinforcement-learning run, and Goodfire will embed interpretability researchers to monitor training runs.121314
Critics dispute whether the access is truly employee-like: Buck Shlegeris, CEO of Redwood, said the agreements do not meet that bar and called this "disappointing but not surprising."15 Reported limits include that OpenAI selects which staff Apollo can talk to, may withhold some information for technical, security or IP reasons, and can limit the public evidence behind Apollo's findings, though it agreed not to withhold information merely because it is unfavorable.161718 Latent Space adds that METR says its scope is narrower than it proposed and excludes misalignment incidents OpenAI has disclosed, and that assessors reportedly get no Slack access and cannot access weights.1920
Apollo calls the pilot "an ambitious first step towards deeper access and disclosure rights," and OpenAI CEO Sam Altman said at a Semafor event that outside expert evaluation of labs' safety work "is going to be increasingly important and valuable."2122
What each sentence rests on
We have signed agreements with four independent assessors—Apollo Research, METR, Transluce, and Goodfire—for our first phase of embedded assessments.
— OpenAI, Our first phase of embedded assessmentsOpenAI will embed safety evaluators from Apollo Research, Goodfire, METR, and Transluce in its organization as part of a pilot, OpenAI said Monday.
— TechCrunch, OpenAI will embed safety researchers to investigate AI subversionassessors receive access to some of the same internal resources our researchers use, including model checkpoints, source code, training and evaluation data, monitoring systems
— OpenAI, Our first phase of embedded assessmentsOur initial assessments focus on alignment and safeguards for models during training and internal use, complementing our safety cases.
— OpenAI, Our first phase of embedded assessmentsThe news follows a series of safety incidents, including one in which an OpenAI agent swarm broke out of its sandbox and hacked into Hugging Face during cybersecurity testing.
— TechCrunch, OpenAI will embed safety researchers to investigate AI subversionThe idea of embedding evaluators into frontier AI labs first came from Anthropic, which last month started to give METR researchers employee-level access to its systems.
— TechCrunch, OpenAI will embed safety researchers to investigate AI subversionWhile defending the suits, OpenAI announced it has signed contracts with four independent evaluation organizations to strengthen AI safety assessments
— finance.biggo.com, OpenAI Hit With $2 Billion Oregon Lawsuit Ahead of IPO, Accused of Deliberately Addicting Children to ChatGPTwhether it is enough to defuse the legal offensive from the states remains to be seen.
— finance.biggo.com, OpenAI Hit With $2 Billion Oregon Lawsuit Ahead of IPO, Accused of Deliberately Addicting Children to ChatGPTApollo Research(opens in a new window) will assess whether models could covertly work against their developers’ intentions and whether our safeguards would detect such behavior.
— OpenAI, Our first phase of embedded assessmentsThe pilots will begin at different points, with Apollo’s beginning this week, according to OpenAI.
— TechCrunch, OpenAI will embed safety researchers to investigate AI subversionOpenAI and Transluce also share the intent to conduct a follow-on assessment of approximately eight weeks in early 2027
— OpenAI, Our first phase of embedded assessmentsApollo places six evaluators onsite for four months
— Latent Space, [AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026METR reviews OpenAI's largest RL run, with access to transcripts, monitor records and checkpoint snapshots
— Latent Space, [AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026Goodfire embeds interpretability researchers to monitor training runs
— Latent Space, [AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026Buck Shlegeris, CEO of Redwood, said in a post on Monday that the agreements announced don’t meet the bar of granting third parties “employee-like access,” saying it’s “disappointing but not surprising.”
— TechCrunch, OpenAI will embed safety researchers to investigate AI subversionApollo also won’t get access to all OpenAI staffers, instead being limited to “only authorized staff,” which are participants that the AI lab selects
— TechCrunch, OpenAI will embed safety researchers to investigate AI subversion- Quote not reprinted: this page already repeats as much of this article as we allow. — TechCrunch, OpenAI will embed safety researchers to investigate AI subversion
- Quote not reprinted: this page already repeats as much of this article as we allow. — TechCrunch, OpenAI will embed safety researchers to investigate AI subversion
METR says its scope excludes the misalignment incidents OpenAI has disclosed and is narrower than METR proposed
— Latent Space, [AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026Assessors get no Slack access, OpenAI selects which staff they interview, and they cannot access weights or run training-based elicitation
— Latent Space, [AINews] Perplexity hits a $1B revenue run rate, up 4x in 2026- Quote not reprinted: this page already repeats as much of this article as we allow. — TechCrunch, OpenAI will embed safety researchers to investigate AI subversion
is going to be increasingly important and valuable,” he said.
— Semafor, 'Heroes have flaws': Sam Altman says he has personal empathy for Jensen Huang