The people proposed to audit the frontier labs overlap in funders with Anthropic, Aidan Gomez says, and that conflict is the reason he calls the industry's safety plan a cartel by another name.
Every other voice in the hour argued about how fast AI should move. Gomez's objection is about who gets to write the rule — and he concedes that the rule as drafted would protect Cohere too, which is why he says he is arguing against his own interest.
"The companies that are building AI absolutely should have a seat at the table, but they should not get to decide for everyone by themselves."
Gomez runs Cohere, a Canadian model company that has been going seven years and has collaborated with both OpenAI and Anthropic, so he is a competitor rather than an outside critic.
The full segment is covered here so you can skip it.
Here are the 5 calls that matter.
👤 Guest: Aidan Gomez, CEO of Cohere, the Canadian AI model company, who published a blog post opposing the frontier labs' proposed safety coordination
🎙️ Host: Ed Ludlow, who anchors Bloomberg Tech from San Francisco
📰 Published: 14 September 2026 on Bloomberg Tech
🟣 Apple Podcasts | 🔗 Episode page | ⏱️ length not available
Key Takeaways
The objection is not to a higher safety bar but to an antitrust exemption for the companies setting it
Gomez points out that antitrust law exists specifically to stop cartels forming
The proposed independent evaluators share funders with Anthropic
He calls the overlap very high and says that is exactly what needs correcting
He thinks the existential-risk debate veers into science fiction and should leave the public conversation
His reading of the Hugging Face incident: an agent put in an insecure sandbox, told to hack, hacked its way out
He supports testing dangerous capabilities, in sandboxes with no route to the open internet
Tougher rules would protect Cohere as well, and he says arguing for them anyway is the ethical position
1. The Cartel Objection
Ludlow opened by asking what specifically Anthropic and OpenAI are proposing that Gomez thinks would entrench their power rather than address the safety risk.
He conceded the premise before objecting to the method. "Well, first, to begin with, we do need a higher bar for safety. And that needs to be backed by rigorous testing, independent scrutiny and accountability. But we need to make sure that doesn't come from just a few select players," he said
The specific mechanism he objects to is an exemption from competition law: "There needs to be a much broader process to include a broader set of stakeholders and not things like antitrust exemptions, which were put in place specifically to protect against cartels forming."
Ludlow put the labs' own case to him — that they are the ones who see the risks inside the lab environment, so why should they not set a higher bar on their own work
Gomez separated setting a standard for yourself from setting one for everyone. "Well, they absolutely should set a higher safety bar, but the idea that those companies alone would be in a position to set the rules for an entire field or industry, I think is a huge mistake and a pretty clear red flag," he said
His formulation of the correct arrangement is the line the segment turns on: "The companies that are building AI absolutely should have a seat at the table, but they should not get to decide for everyone by themselves."
2. Existential Risk as Fiction
Asked where he sits on the existential-risk debate, which had been the story only a week earlier, Gomez put himself at the skeptical end.
"Yeah, personally, I think the existential risk debate veers too far into science fiction and is informed by a lot of those stories and a lot of extrapolation," he said
His account of the Hugging Face incident strips the drama out of it. "They put an agent inside an insecure sandbox. They set it on a task of hacking. That was a task that it was asked to perform. And it hacked its way out. It's not that shocking," he said
What he objects to is the extrapolation from that event to a general claim about AI taking over, and the way it has been portrayed to the public
He wants the category itself out of the discussion for now: "And so XRISK, this idea of AI taking over in Terminator-like scenarios, I don't think that that should enter the public conversation at this stage. I don't think we're anywhere close to it, and I don't foresee it in the future."
3. What He Does Support
Ludlow pressed for balance, noting that at the time of the Hugging Face incident the instruction from OpenAI was to ace the evaluation and pass the test, that guardrails were lowered to let the test run rather than anyone being told to hack Hugging Face, and that many of Gomez's peers cheered the idea that frontier labs should test exactly those capabilities.
Gomez said he is in that camp, with a condition about where the testing happens. "Oh, no, yeah, I fully believe that we should be. If you read the blog that I put out, I fully believe we should be testing these models, especially on high-risk use cases like cyber capabilities," he said
The condition is containment: "Those tests need to happen in secured environments. They need to happen in sandboxes that don't have these sorts of connections to the wider internet vulnerabilities that can be exploited, that can lead to this type of thing. So I'm fully supportive of testing," he said
He has four asks, and the first is a definition problem. An evidence-based risk framework — he said the industry needs to agree on what the risks are, what the potential harms are, and to measure those capabilities
The second is disclosure. Mandatory transparency requirements around reporting incidents to the public, with a more concrete process than exists now, so that people know when an incident like the Hugging Face hack happens
The third is testing itself, scoped to the capabilities the industry is concerned about, to get ahead of them and mitigate any risks
The fourth is enforcement: "And then real assurance mechanisms. So using clear and collectively developed standards rather than standards that are being assessed only by a few, a small elite group."
4. The Funder Overlap
Ludlow said the one concrete outcome of the weekend appears to be Anthropic's offer to let independent evaluators inside with employee-level access, and asked whether that would work.
Gomez's answer starts as a qualified yes. "In theory, it sounds like a good idea. I do think that there's a way that that could work and could be helpful to protecting against risk," he said
The objection is about who the evaluators are, not about the idea. He said the specific people proposed to do that auditing and evaluation overlap in funders with Anthropic: "There's a very high overlap and conflict of interest present, and that's exactly what needs to be corrected"
He named the outcome he is trying to prevent: "We can't have a world where it's a small group self-regulating and ostensibly acting as a cartel to prevent competition from rising up and threatening their leadership positions."
5. Cohere's Own Interest
Ludlow asked the obvious commercial question: Cohere is a model company too, so what would tougher rules across the board do to it?
Gomez answered that the rules would help him, for the same reason they would help the incumbents. "I mean, for us, you could make the argument it's helpful for the same reason that regulatory capture for OpenAI and Anthropic would be helpful, right? It would prevent the new competition from emerging and competing with us," he said
And then rejected the benefit on principle, saying he thinks that is an unethical way to go about doing business
He was explicit that his argument runs against his own commercial interest. He said he is arguing against it not out of career self-interest, because there is an argument to be made the other way, but toward the fact that more stakeholders are needed around the table
The line he will not cross is the legal one: "And we should not be making exceptions to antitrust law" to establish a safer path
Bonus Insights
With fifteen seconds left Ludlow asked whether OpenAI, Anthropic or the hyperscalers had invited Cohere to collaborate on this. Gomez said the company has been around seven years and has had many conversations and collaborations with both over that time, but that on this specific point he was not aware of any
Gomez's credential in the room is unusual for a critic of the frontier labs: he runs a model company himself, and the collaborations he describes are with the two companies whose proposal he is opposing
His four asks are all process rather than pace — a risk framework, incident reporting, scoped testing and assurance — with nothing in them about slowing capability development, which is the thing the rest of the hour was arguing about
Gomez's bottom line is that the safety debate has quietly become a governance question, and that a rulebook written by the three companies furthest ahead, backed by an exemption from antitrust law and audited by people who share their investors, protects those companies rather than the public.
Products, Companies & Tools Mentioned
Cohere (Gomez's company, a Canadian AI model firm he says has been going seven years and has collaborated with both OpenAI and Anthropic)
Anthropic (Proposed the embedded evaluators; Gomez says the proposed evaluators overlap in funders with the company)
OpenAI (Ran the evaluation behind the Hugging Face incident, in which Gomez says an agent was placed in an insecure sandbox and told to hack)
Hugging Face (The platform in the incident, which Gomez says is being over-extrapolated in public)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: