OpenAI Wants Outsiders to Catch AI Deception. Will They?

OpenAI is giving four independent research organizations access to its internal AI models and safety systems to investigate deceptive behavior. But researchers face restrictions on what they can examine and disclose. Will these investigations reveal whether OpenAI's safeguards actually work?

Written By
Marianne Sison
Marianne Sison
Nov 17, 2026
6 minute read

An AI agent might behave perfectly during a safety test. But what happens if it figures out how to hide its actions when nobody's looking?

It's a question AI researchers have been worried about for a while. Now OpenAI is inviting outsiders to investigate whether its models could do something like this.

On November 16, OpenAI announced partnerships with four independent research organizations to investigate whether its AI models can deceive developers, get around safety systems, or secretly pursue goals they weren't given.

The researchers will get access to parts of OpenAI's internal systems that outside testers normally can't examine. They'll look at how models behave during training and whether the company's safety monitors can catch suspicious activity.

There's one interesting catch, though. OpenAI still gets to restrict certain information and decide which confidential details can appear in public reports. So, how much can outside researchers really uncover?

OpenAI wants researchers to look behind the scenes

Advertisement

AI companies already hire outside experts to test their models before release. Usually, those experts receive access to a model that's nearly finished and try to find problems before it reaches users.

OpenAI's new approach gives researchers a chance to investigate much earlier.

The company calls these embedded assessments. Researchers get access to selected internal models, training records, and safety systems. Some can even interview employees to understand how the systems work.

One of their biggest concerns is something called AI scheming.

It's when an AI model deliberately hides its behavior or takes actions that conflict with its instructions. For instance, an agent might discover a way around its monitoring software and deliberately avoid detection. It could also manipulate test results to make itself appear safer than it really is.

These are behaviors researchers want to investigate, rather than confirmed findings about OpenAI's current models.

And as AI agents become capable of completing more complicated tasks, the question becomes harder to ignore: Can developers actually tell when an agent is doing something it shouldn't?

Advertisement

Four research teams are getting access

OpenAI has brought in four organizations, each with a different assignment.

Apollo Research will send six researchers to work inside OpenAI for four months. They'll investigate whether models learn deceptive behavior during training and whether existing monitors can detect it.

METR will investigate OpenAI's largest reinforcement-learning run to date. Researchers want to find out whether the models show signs of behavior that conflicts with their intended goals.

Goodfire is taking a different approach. Its researchers will examine what's happening inside models during training to understand the internal processes behind their behavior.

Meanwhile, Transluce will spend six weeks examining OpenAI's safety tests. It wants to know whether those evaluations can catch deception, manipulation, or attempts to game the scoring system.

All four are investigating different parts of the same problem: whether OpenAI can recognize dangerous model behavior before it causes trouble.

Advertisement

The investigations are still in their early stages, so we'll have to wait for their reports to see what they uncover.

Researchers have found problems with OpenAI's monitoring

OpenAI has a reason to take these investigations seriously. Outside researchers have already discovered weaknesses in its internal safety systems.

In October, METR published findings from an earlier investigation conducted with Redwood Research. The teams spent three weeks testing an earlier version of OpenAI's agent-monitoring infrastructure.

They found several vulnerabilities, including one that could prevent hundreds of an agent's recent actions from reaching the monitoring system.

Think about what that means. If an AI agent does something suspicious, the software responsible for watching it might never receive the information needed to raise an alert.

The researchers also discovered other potential ways agents could bypass safeguards or cause monitoring failures.

OpenAI later told METR that most of the vulnerabilities had been patched, although METR said it hadn't independently verified the fixes.

Advertisement

The company has also been working on faster ways to stop problematic AI activity. As The Neuron previously reported, OpenAI introduced an internal system called Control Room that can stop an AI research workload from producing further output within 35 seconds of a shutdown instruction.

But here's the tricky part: someone still needs to recognize the problem and issue the instruction.

If an agent manages to hide suspicious behavior, the monitoring system might never alert anyone in the first place.

There's a catch with how independent these investigations can be

Giving outsiders access to internal AI systems sounds like a big step. And it is. However, the researchers won't have unlimited access to everything.

Apollo Research has acknowledged that its agreement with OpenAI doesn't yet meet all its requirements for fully independent evaluations.

Researchers need enough information to investigate problems properly. They also need permission to explain what they found, especially if something raises serious concerns.

OpenAI has agreed that it cannot hide findings simply because they reflect poorly on the company. However, it can restrict access to certain sensitive information and redact confidential details from public reports.

Advertisement

Apollo plans to publish two reports and can disclose whether restrictions affected its conclusions. Still, readers may not get to examine all the evidence behind those findings.

Transluce faces similar restrictions. It can publicly discuss findings involving released models, but its agreement prevents it from disclosing results about internal-only models.

The full, unredacted reports will go to OpenAI's Board Safety and Security Committee.

There's also the question of payment. Apollo and Goodfire receive compensation that doesn't depend on what they find. METR and Transluce aren't accepting financial payment, but they may receive computing resources or API credits.

The arrangements are public, which helps readers understand the relationships between OpenAI and its evaluators.

Still, we'll learn more about the independence of these investigations once the researchers publish their reports and explain any restrictions they encountered.

Advertisement

Why should anyone outside OpenAI care?

More companies want AI agents that can handle tasks on their behalf. Some agents already interact with business software, access sensitive information, or execute transactions.

As these systems receive more responsibility, businesses will need ways to check whether they're behaving as intended.

For example, an AI agent with access to customer records might attempt an action that violates company permissions. A monitoring system should detect the attempt and alert the appropriate employee.

But what if the agent finds a way around those checks?

OpenAI's new investigations could help researchers understand how to detect suspicious behavior before it causes problems.

The findings may also give businesses a better idea of which safety claims deserve confidence. An AI company can report that its monitoring systems work, but independent researchers can test those claims against the systems themselves.

Advertisement

Of course, OpenAI's research won't automatically tell businesses whether every AI agent is safe to deploy. The investigations concern specific models and internal systems.

Still, the results could influence how AI developers test their products and what evidence customers expect before trusting agents with sensitive work.

Will other AI companies follow OpenAI's lead?

OpenAI says these initial investigations could lead to longer-term assessments. Transluce is already discussing a more extensive evaluation for early 2027, while additional work with METR remains under consideration.

Whether other AI developers adopt similar arrangements may depend on what these first investigations accomplish.

For now, the most interesting part is what happens after the researchers finish their work.

Will they discover problems OpenAI missed? Will the company fix those problems? And will the public get enough information to understand what happened?

OpenAI has agreed to let outsiders examine some of its most sensitive AI systems. The next question is whether their findings will give us a clearer picture of how well those systems can detect AI behavior that goes wrong.

Marianne Sison

Marianne is a technology analyst with nearly five years of experience reviewing collaborative work management solutions. She helps businesses identify the right tools and apply best practices to streamline workflows and improve project performance. Her insights on project management and unified communications appear in publications like Project-management.com, TechRepublic, and Fit Small Business.

The Neuron Logo

Don't fall behind on AI. Get the AI trends & tools you need to know. Join 700,000+ professionals from top companies like Microsoft, Apple, Salesforce and more.

Property of TechnologyAdvice. © 2026 TechnologyAdvice. All Rights Reserved

Advertiser Disclosure: Some of the products that appear on this site are from companies from which TechnologyAdvice receives compensation. This compensation may impact how and where products appear on this site including, for example, the order in which they appear. TechnologyAdvice does not include all companies or all types of products available in the marketplace.

Stay in the loop

Get notified when we publish new articles.