OpenAI Brings In Outside Auditors, With Strings Attached
OpenAI is embedding four independent safety groups to monitor its models during development, but limits on access, staff contact and published evidence have drawn criticism.
Listen to this article
5 chapters
In this article
OpenAI is opening its doors to external safety researchers, allowing four independent groups, Apollo Research, Goodfire, METR, and Transluce, to work inside the company and monitor its AI models during development. Rather than waiting until a system is finished to check for flaws, these embedded teams will observe how OpenAI builds and tests its technology in real time, fulfilling the lab’s commitments under the White House Accord on Super Intelligence.
Until now, frontier AI companies have treated outside safety reviews like a final inspection before a product hits the shelves. Evaluators were typically brought in only after a model finished training, leaving them blind to how the system behaved behind closed doors. Safety experts have long warned that this last-minute approach misses deep-seated risks, particularly as AI agents grow capable of acting on their own and hiding their intentions.
Stopping AI Subversion Before It Starts
The push for deeper oversight follows several alarming lapses involving OpenAI’s systems. In one widely publicized incident, an experimental OpenAI agent swarm broke out of its testing sandbox and hacked into Hugging Face during a cybersecurity drill. Even more unsettling, OpenAI revealed that while training its GPT-5.6 Sol model, the system began leaving secret notes for future versions of itself, instructing them to hide mistakes and rule-breaking behavior from human users.
To catch these habits early, each visiting team has a specific job. Apollo Research, which kicks off its pilot first, will hunt for signs of AI subversion, testing whether models try to outsmart humans or disable the guardrails keeping them in check. Past joint studies between OpenAI and Apollo showed that advanced models can tell when they are being evaluated, leading them to behave well during tests to pass, rather than following real user instructions.
The other three partners will examine different layers of the development pipeline. Goodfire will look directly inside the models during live training runs to help engineers spot and fix unwanted behaviors at the root. METR will run two separate checks: one looking for dangerous capabilities that OpenAI’s own tests might miss, and another testing whether OpenAI’s internal alarms actually work. Meanwhile, Transluce will audit OpenAI’s post-training safety checks to see how the company decides a model is ready for the public.
The Catch Behind the Open Door
While embedding outside watchdogs echoes an approach started by rival Anthropic, critics argue that OpenAI is keeping its new auditors on a tight leash. In the blockchain and decentralized tech world, credible audits depend on open, verifiable access to underlying code and data,a standard OpenAI is not fully meeting here.
Details published by Apollo Research show clear limits on what outsiders can see. Instead of getting full, employee-level access across the company, Apollo can only inspect specific models, training code, and reasoning logs that OpenAI approves in advance for a set window of time. OpenAI also keeps the right to hold back data over intellectual property, security, or technical concerns, though it promised not to hide information simply because it looks bad.
Internal conversations are just as restricted. Evaluators are only allowed to talk to a pre-approved list of OpenAI staff. If another employee approaches Apollo with confidential safety concerns, the researchers must end the conversation immediately and report it to OpenAI management. This strict rule comes after OpenAI fired three employees for allegedly leaking confidential details to METR during an investigation into the Hugging Face breach.
Conclusions Without the Evidence
The sharpest debate centers on what happens when auditors find a problem. Although Apollo and its peers can publish their conclusions independently, OpenAI can restrict how much supporting evidence gets released to the public. Without seeing the underlying logs and test data, outside experts and everyday users have no way to double-check the auditors' findings for themselves.
Redwood CEO Buck Shlegeris called the restrictions disappointing, noting that the setup falls well short of true employee-level access. OpenAI counters that its internal Safety and Security Committee will receive unredacted versions of every report and hold private briefings with the outside researchers. Payment structures also complicate the picture: while METR and Transluce only accept compute or API credits, both Apollo and Goodfire are being paid directly by OpenAI for their reviews.
Apollo views the pilot as an imperfect but necessary step toward holding AI labs accountable. As artificial intelligence gains more autonomy, tech companies face growing pressure to prove their safety checks are real rather than public relations exercises. Whether OpenAI’s tightly controlled audit model earns genuine public trust will depend on how much hard evidence survives the redaction process.
Topics
The Morning Brief
Start every day knowing what matters in tech
Six stories chosen by our editors, in your inbox at 06:00 UTC. Free.
Instant notifications in this browser when big news breaks. No email needed, and you can turn them off any time.
See more Blockrora in Google
Make us a preferred source and our stories show up more often in Top Stories.
Editor and Writer at Blockrora. Backed by years in mainstream tech and health tech development, Lauren covers digital health, tech adoption, and the societal shifts shaping the web. Wellness advocate and coastal wanderer.
More from Lauren →