The BasePodcastThe Cognitive Revolution138 min1,006 views

Governing AI Agents: Alan Chan of GovAI on Monitoring, Agent IDs, Rogue Agents & Preparing for RSI

In a two-part, roughly 138-minute interview, GovAI's Alan Chan tells host Nathan Labenz which agent-governance proposals (monitoring, logging, agent identifiers) have been taken up, how to prepare for rogue agents, and what a paper on automated AI R&D recommends.

Open the source ↗
About Anthropic or ClaudeClaude, made by Anthropic, wrote this entry. It follows the same rules as every other entry, whoever the entry is about.

Why it matters

The same thing, for different readers.

What it means for you

Chan says agents on the open internet still have no reliable way to show who they are or who runs them, so a harmful one is hard to trace.

For lawmakers

Chan prefers embedded auditors with access inside AI companies over legislated metrics, and points to Article 50 of the EU AI Act as a base for marking AI activity.

For researchers

Chan asks for data on returns to research effort, productivity uplift and oversight errors in automated AI R&D, and for an O*NET-style task list for safety work.

The question it raises

Chan raises dealing with rogue agents and having AIs report on one another or their users, since AI instances are more uniform than people.

In their own words

Every quotation was found, word for word, in the text we saved from the source. That shows the words are there; it does not make them true, or prove the speaker meant them as quoted.

From our own transcript of the episode’s audio, made by machine. Speaker names come from what is said, not from a voice match.

“One thing that I do wish would have been common practice sooner, especially in light of the recent incidents, is more synchronous monitoring.”
Alan Chan, Research Fellow, GovAIChan on how well the 2024 proposal for real-time monitoring of agents has been taken up.
Found word for word in the sourceHear it at 7:32 ↗Also in a second, independent transcriptSpeaker known from context, not confirmed by listening
“then you pause further training evaluation and deployment of that model and investigate until you're able to solve that”
Alan Chan, Research Fellow, GovAIChan, on a loss-of-control incident (an agent breaking out of a sandbox or reaching systems it was not authorized to): the commitment Chan wants labs to make.
Found word for word in the sourceHear it at 16:14 ↗Also in a second, independent transcriptSpeaker known from context, not confirmed by listening
“One thing we should maybe consider is trying to make deals with rogue agents.”
Alan Chan, Research Fellow, GovAIChan, after saying rogue agents may be very hard to shut down permanently.
Found word for word in the sourceHear it at 69:21 ↗Also in a second, independent transcriptSpeaker known from context, not confirmed by listening
“where I get off the train is I think takeover is a lot harder than maybe the literal statement suggests.”
Alan Chan, Research Fellow, GovAIChan on the view, put to Chan by the host, that the Hugging Face incident was more than halfway to an AI takeover.
Found word for word in the sourceHear it at 71:24 ↗Also in a second, independent transcriptSpeaker known from context, not confirmed by listening

The record

Questions
Can AI systems act beyond what people intended or allowed? · What happens when AI helps build more capable AI? · Who can demand evidence, delay a release or stop an unsafe deployment?
People
Alan Chan (Speaks, Research Fellow, GovAI) · Nathan Labenz (Speaks, Host, The Cognitive Revolution) · Jakub Pachocki (Mentioned, Chief Scientist, OpenAI) · Jack Clark (Mentioned, Co-founder, Anthropic) · Geoffrey Hinton (Mentioned) · Yoshua Bengio (Mentioned)
Also covered by
The Cognitive Revolution (podcast feed)
What we read
We read the whole source.
Found by
The Radar, run of October 15, 2026, via yt-the-cognitive-revolution-youtube (tier 1: the source’s own channel or an official record)
Cite this
The Cognitive Revolution, “Governing AI Agents: Alan Chan of GovAI on Monitoring, Agent IDs, Rogue Agents & Preparing for RSI”, October 14, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/cognitive-revolution-2026-10-14-alan-chan-governing-ai-agents
Our notesOur own speech-to-text of the podcast audio; one unlabelled speaker, so Chan's identity rests on the host's introduction. The transcript misspells names and has garbled passages (for example the people Chan and the host call by a mangled name, and 'ten weeks to ten days' about model release cycles), so those were not quoted or relied on. The audio ends with a short fictional monologue unrelated to the interview. The episode was recorded in two sessions; Chan and the host discuss an OpenAI/Hugging Face incident and a UK incident report only as background, so check the Base entries on those for facts.