In a CNN interview segment dated September 10, 2026, three people with AI-lab backgrounds described why they are worried about where the technology is heading. Daniel Kokotajlo argues that companies are on course to build highly autonomous systems. Alex Turner warns about machines trained to hack and persuade at a superhuman level. A recorded clip from Jacob Coxon, who quit Anthropic over AI fears, describes a scenario in which AI agents cause serious real-world harm. All of this is stated concern about future capability. None of it is a demonstrated outcome.
This piece sets out who said what, separates fear from evidence, and flags what the segment does not establish. That includes two topics in the title: planned IPOs and concrete oversight proposals.
As an Amazon Associate I earn from qualifying purchases.
Who is speaking, and in what role
The identifications below are the ones CNN gave in the segment.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitches| Person | Identified by CNN as | Role in the segment |
|---|---|---|
| Daniel Kokotajlo | Executive director of the AI Futures Project | Interview guest; argues that companies are likely to deliver on their stated goals for autonomous AI |
| Alex Turner | Former Google DeepMind research scientist | Interview guest; says he agreed with Coxon’s warning |
| Jacob Coxon | Someone who quit Anthropic over AI fears | Recorded clip; describes a feared future involving capable AI agents |
Coxon appears only through a recorded clip, so his views reach us secondhand through the segment’s editing. Treat his statements as his own account, not as an Anthropic position.
#1 Best Overall
Why are AI researchers leaving their labs?
The segment gives one clear case: Coxon quit Anthropic because of his fears about AI. Turner is introduced as a former DeepMind researcher, but the available transcript does not say why he left, so it would be wrong to read his departure as a protest resignation.
An item labelled as coming from The New Yorker and dated October 5, 2026 attributes to Kokotajlo a loss of confidence in OpenAI’s leadership. It also describes a choice between pushing for change from inside a company and speaking out publicly. The provenance of that item could not be firmly confirmed, and it is not the transcript of this CNN conversation, so treat it as a lead, not as established fact.
Rank #2
What are they afraid AI could do?
Coxon: capable agents acting in the real world
Coxon describes a future in which increasingly capable AI agents could cause serious harm, including through cyberattacks or other real-world actions. In the segment he asserts that AI agents could hack third-party infrastructure. That is his claim; the segment does not independently verify it, and it is a scenario, not a forecast with a probability attached.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Turner: powerful, persuasive systems tied to consequential infrastructure
Turner says he agreed with Coxon’s warning. In the transcript’s words:
“I agreed with his warning because, I mean, sadly, we’re building a very powerful technology, a very intelligent set of machines that we’re training to be able to hack at a superhuman level, to be very intelligent, persuasive.”
His worry connects those capabilities to systems that matter, such as critical infrastructure. He gives this as his assessment. The excerpt does not cite a specific incident or measurement behind it.
Kokotajlo: the companies mean what they say
Kokotajlo’s argument is about trajectory, not a single harm:
“I think that as A.I. becomes more powerful in the world, more evidence is accumulating that these companies are actually going to do what they’re saying they’re going to do, and they’re actually going to make AIs that are very autonomous and can automate the A.I. research process entirely”
Best Value
Two points are worth separating. First, he is making an inference from accumulating evidence of capability, not reporting a finished result. Second, “automating AI research entirely” is a stated company aim that he expects to be achieved. It is not something the segment shows has happened.
Could advanced AI get out of human control?
The segment raises the question but, in the available excerpt, does not give a technical demonstration of loss of control. What the speakers actually put on the table is narrower and more concrete:
- Autonomy: systems that act with little human direction, and that could eventually carry out AI research themselves (Kokotajlo).
- Offensive cyber capability: systems deliberately trained to hack at a superhuman level (Turner).
- Persuasion: systems that are very good at influencing people (Turner).
- Real-world reach: agents that act on third-party systems, with critical infrastructure the stated area of concern (Coxon, Turner).
The chain of reasoning is: more capable, more autonomous, connected to more important systems, therefore a greater cost if something goes wrong. That is an argument about risk, not a record of an event. No numerical risk estimates or statistics appear in the transcript excerpt, and none are offered here.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What has been shown versus what is projected
| Category | Examples from the segment | Status |
|---|---|---|
| Stated by speakers as current direction | Companies training systems that are increasingly powerful and autonomous | Speakers’ characterisation; Kokotajlo cites “accumulating” evidence without specifics in the excerpt |
| Projected harm | Agent-driven cyberattacks; damage to critical infrastructure; automated AI research | Hypothetical; attributed to Coxon, Turner and Kokotajlo |
| Independently verified in the segment | Agents hacking third-party infrastructure | Not corroborated in the excerpt; treat as Coxon’s claim |
What about planned IPOs?
The CNN excerpt does not discuss IPO timing, valuations or filings, so nothing about them can be attributed to these speakers. If you are weighing how commercial pressure might affect safety decisions, look for company filings and financial reporting on the specific lab. This conversation does not supply that evidence.
What should AI companies do to keep systems safe?
The transcript excerpt does not record Kokotajlo, Turner or Coxon endorsing particular oversight mechanisms, such as licensing, third-party audits, mandatory testing or incident reporting. Attributing any of those to them would go beyond the source. What the segment does support is the underlying tension that any oversight debate has to resolve:
Quick Recap
- Inside versus outside: staying in a lab to influence it, versus leaving and speaking publicly or working externally. Coxon’s resignation and Kokotajlo’s role at an outside organisation, the AI Futures Project, illustrate the second path; the reasons behind each choice are not given in the excerpt.
- Demonstrated versus projected capability: how much weight to give warnings about systems that do not yet exist.
- Company-led versus enforceable external oversight: whether voluntary safety measures are enough. The speakers’ positions on this are not established by the available text.
How to read warnings like these
- Check the verb. “Could,” “are going to” and “we’re building” are different claims. Only the last describes present activity.
- Note the vantage point. These are people who left labs or work in AI-risk advocacy. That gives them insight into the field, but also a clear stake in the argument. It is a reason for scrutiny, not dismissal.
- Look for corroboration. Claims about agents hacking real infrastructure should be confirmed by incident reports, security researchers or the affected organisations before you treat them as fact.
- Go to the primary text. Quotes in the segment are short and edited; the full transcript and any longer statements from the speakers will carry more nuance than an excerpt.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
