Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarTNW→ original

OpenAI Lets Six Apollo Auditors Inside for Four Months

OpenAI has let six auditors from Apollo Research inside for four months to look for signs that its models are circumventing safety constraints. Until now, Apollo only received models shortly before release. This time it will see training, testing, and internal use. Access is limited: OpenAI chooses who the auditors can talk to and can withhold some materials. The arrangement is a test of whether external AI auditing is possible without censored conclusions.

AI-processed from TNW; edited by Hamidun News
OpenAI Lets Six Apollo Auditors Inside for Four Months
Source: TNW. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTNW ↗
Editor-in-Chief
Key metric #1Apollo will publish two reports; OpenAI cannot remove Apollo's verdicts
Key metric #2OpenAI is providing training code, data, and reasoning traces from training runs
Key metric #3No access to Slack; work is limited to OpenAI-provided hardware

Normally, external AI safety testing looks like this: a lab finishes a model, hands it over for testing a few weeks before release, and then publishes a reassuring report. Apollo Research and OpenAI have decided to try a different format. Six Apollo evaluators will spend four months inside OpenAI, checking whether models attempt to undermine the control and alignment measures meant to keep them in bounds.

What exactly will be checked

The review covers not only the final model but the entire pipeline: training, testing, and internal use. This matters because, in Apollo's view, it is too late to catch a situation just before release in which a model quietly pursues goals its creators never built in. By that point, such behavior could already have become entrenched during training.

The team will look for answers to four questions. First: does training reward the model for circumventing constraints, and can this be detected with tests? Second: does this behavior appear when OpenAI itself uses its models internally, and will the model give itself away through its reasoning? Internal use is a real risk: in July, OpenAI already suspended one of its internal models.

For the first area, training, OpenAI will provide training code, data, and reasoning records from actual training runs. Access for the other three areas has not yet been agreed.

Where the boundaries lie

The conditions fall well short of full freedom. OpenAI itself chooses which employees the evaluators can speak with, and official conversations are logged. The team will work on OpenAI's hardware, through channels that can be monitored, and without access to the corporate Slack, where much of the informal discussion happens.

OpenAI may hold back materials for safety or intellectual property reasons, but not because they are inconvenient. Apollo will publish two reports. OpenAI may ask for confidential evidence to be redacted, but it cannot remove Apollo's conclusions themselves. Moreover, Apollo may publicly describe what it was denied and how that affects the reliability of its findings.

This last point is the most interesting. Apollo calls it meta-transparency: if the auditors find nothing, that does not mean everything is clean. They may simply not have been shown where the problems are. Disclosing the constraints lets readers judge for themselves how much a "clean" result actually proves.

Context and what to do about it

The deal continues OpenAI's recent moves: Sam Altman has previously spoken about external testers with access close to that of employees, and the company planned to let third-party groups test models during training itself. Against this backdrop, Apollo calls the pilot only a first step: full access and the freedom to publish the results it is seeking are still a long way off.

A practical takeaway for developers and product managers building on OpenAI models or their own agent systems. First, an "auditor inside the process" model may become standard for large labs, and customers will start asking how model behavior is verified. Second, when reading any safety report, look not only at the results but at the conditions: what the auditor was shown, what was withheld, and who chose the interviewees. If you are setting up internal auditing for your own agents, adopt the same principle: define in advance what cannot be edited out of the conclusions, and openly describe the access limitations.

An honest caveat: a four-month pilot with six people does not guarantee that dangerous behavior will be found, and some access has not yet been agreed. But the fact that the auditors' verdicts cannot be struck out, and that refusals can be made public, makes this experiment considerably more substantive than a typical pre-release check.

Story

Key Questions Answered

What exactly will be checked?

The review covers not only the final model but the entire pipeline: training, testing, and internal use. This matters because, in Apollo's view, it is too late to catch a situation just before release in which a model quietly pursues goals its creators never built in. By that point, such behavior could already have become entrenched during training. The team will look for answers to four questions. First: does training reward the model for circumventing constraints, and can this be detected with tests?

Where the boundaries lie?

The conditions fall well short of full freedom. OpenAI itself chooses which employees the evaluators can speak with, and official conversations are logged. The team will work on OpenAI's hardware, through channels that can be monitored, and without access to the corporate Slack, where much of the informal discussion happens. OpenAI may hold back materials for safety or intellectual property reasons, but not because they are inconvenient. Apollo will publish two reports.

What are the key details on Context and what to do about it?

The deal continues OpenAI's recent moves: Sam Altman has previously spoken about external testers with access close to that of employees, and the company planned to let third-party groups test models during training itself. Against this backdrop, Apollo calls the pilot only a first step: full access and the freedom to publish the results it is seeking are still a long way off. A practical takeaway for developers and product managers building on OpenAI models or their own agent systems.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…