People who build AI

Mark Chen

Chief Research Officer at OpenAI

4 records

Records

Yeah, I think so. And you know, today is a world where we’re moving towards AI with connectors. So these are AI models that can plug in to your email, your Google Docs, your Slack. And I think that poses real risks. People today are fairly good at jailbreaking models, too, if there’s a very motivated hacker. So what is the implication, right?

Exact quote, beginning

Mark Chen

Chief Research Officer at OpenAI

Mark Chen said AI systems connected to email, Google Docs, and Slack could let a motivated attacker take information or launch a coordinated attack.

Chief Research Officer, OpenAIPublished June 25, 2025 · date of remarks not establishedTech Unheard podcast, Episode 7

Source Archived

Context and checks

Surrounding words

Rene Haas[13:26] Yeah. I don’t know how much you interface with potential clients, but do you hear the safety thing increasing, now, given with the capability of certainly [GPT-]4.5 and [OpenAI] o3? Mark Chen[13:35] Yeah, I think so. And you know, today is a world where we’re moving towards AI with connectors. So these are AI models that can plug in to your email, your Google Docs, your Slack. And I think that poses real risks. People today are fairly good at jailbreaking models, too, if there’s a very motivated hacker. So what is the implication, right? You could imagine that someone motivated extracts all that information away from you or they’re able to launch some kind of coordinated attack. Rene Haas[14:03] Right, so then literally at the source code level, you could just simply put things inside the model that when those queries are requested, be rejected. Mark Chen[14:12] Right. Yeah.

What this quote does not say

The transcript says 'some kind of coordinated attack' without specifying the target, scale, or mechanism.
Later in the interview, Chen described connected personal agents as desirable; he did not withdraw the connector risks stated here.

How this was checked

Quote matched character-for-character against the captured page.

Reviewed by

an AI reviewer that read the whole source · an automatic character-by-character check

CoTs can look benign if the malign reasoning can be done in activations, so a model can be misaligned without visible malign reasoning. Care must be taken not to create a false sense of safety based on such monitoring.

In a 2025 paper, Mark Chen and his co-authors recommended monitoring an AI model's written reasoning only as an added safety measure, saying dangerous reasoning could remain hidden and create false confidence.

Chief Research Officer, OpenAIPublished December 7, 2025 · date of remarks not establishedChain of Thought Monitorability: A New and Fragile Opportunity for AI Safety

Source Archived

Context and checks

Surrounding words

In cases where thinking out loud is not required, CoT Monitoring can detect some misbehavior, but does not by itself produce strong evidence of safety. CoTs can look benign if the malign reasoning can be done in activations, so a model can be misaligned without visible malign reasoning. Care must be taken not to create a false sense of safety based on such monitoring. Monitoring for dangerous tasks that need reasoning may not catch all relevant harms.

What this quote does not say

The paper later recommends chain-of-thought monitoring only as an addition to other safety research, not as a replacement; this passage qualifies rather than rejects the method.
The paper has 41 authors; authorship supports endorsement of the paper but does not establish that Chen drafted this sentence.

How this was checked

Quote matched character-for-character against the captured page.

Reviewed by

an AI reviewer that read the whole source · an automatic character-by-character check

Developers should consider measures of monitorability alongside other capability and safety evaluations when deciding to train or deploy a given model. These decisions should then be based on holistic assessments of risk which account for CoT monitorability, performance characteristics of monitoring systems, and estimates of models’ propensity for misbehavior.

Mark Chen

Chief Research Officer at OpenAI

In a 2025 paper, Mark Chen and his co-authors said AI developers should consider how well a model's reasoning can be monitored, along with other safety tests, when deciding whether to train or deploy it.

Chief Research Officer, OpenAIPublished December 7, 2025 · date of remarks not establishedChain of Thought Monitorability: A New and Fragile Opportunity for AI Safety

Source Archived

Context and checks

Surrounding words

Use monitorability scores in training and deployment decisions. Developers should consider measures of monitorability alongside other capability and safety evaluations when deciding to train or deploy a given model. These decisions should then be based on holistic assessments of risk which account for CoT monitorability, performance characteristics of monitoring systems, and estimates of models’ propensity for misbehavior. For example: - (a) Developers might consider whether to proceed with a novel model architecture that does not have monitorable CoT and then document their decision in the system card if the model is deployed; - (b) If

What this quote does not say

The paper has 41 authors; authorship supports endorsement of the paper but does not establish that Chen drafted this sentence.
This is a technical developer recommendation, not a government policy position.

How this was checked

Quote matched character-for-character against the captured page.

Reviewed by

an AI reviewer that read the whole source · an automatic character-by-character check

Statements they signed

FROM THE SIGNED STATEMENT

It is hard to predict exactly how much this will accelerate AI progress, but there is a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems.

Signed statement: “Pacing the Frontier” (frontier-lab employees, Guidelight + Encode AI)

Mark Chen

Chief Research Officer at OpenAI

Signed as part of a group statement. Not spoken.

Mark Chen signed Pacing the Frontier, which asks the U.S. government to support an international effort to develop tools for deliberately slowing automated AI development when needed.

Chief Research Officer, OpenAIDate not establishedPacing the Frontier

Source Archived

Context and checks

Surrounding words

Scientist, OpenAI Jared Kaplan Co-Founder and Chief Science Officer, Anthropic Shengjia Zhao Chief Scientist, Meta AI Shane Legg Co-Founder & Chief AGI Scientist, Google DeepMind Ilya Sutskever CEO, Safe Superintelligence Inc. Mark Chen Chief Research Officer, OpenAI Jasjeet Sekhon Chief Strategy Officer, Google DeepMind Dario Amodei CEO, Anthropic Jack Clark Co-Founder and Head of Public Benefit, Anthropic Anca Dragan VP, AI Safety & Alignment, Google Wojciech

What this quote does not say

The role is copied from the roster's wording: “Chief Research Officer, OpenAI.”

How this was checked

The organiser's own roster lists this person as a signer.

Reviewed by

an AI reviewer that read the whole source