# The insiders went public. OpenAI fired three, then Anthropic wrote down the rules

URL: https://www.thedeepfeed.ai/posts/2026-10-17-the-people-leaving-the-labs/
Category: People
Published: 2026-10-17
Author: the-deep-feed
Tags: people, openai, anthropic, safety, whistleblowers, metr, insider-risk, evaluators
Kind: deep

> Between September 9 and October 15, a resignation thread, an Atlantic essay, three fired safety researchers and an arrest made OpenAI and Anthropic staff the public's main interpreters of the labs, and the two companies answered their own people in very different ways.

## TL;DR

- **175 million** impressions: Jacob Coxon's Sep 9 resignation thread said neither OpenAI nor Anthropic 'is acting responsibly.' The heavier signal came **83 minutes** later, when Anthropic's alignment science lead **Evan Hubinger**, who stayed, replied 'Jacob is correct here.'
- OpenAI fired **three** safety researchers over what it calls a 'significant breach of trust' it has not described. Tomek Korbak says the only reason he was given, verbally, was how he communicated with **METR**, the outside auditor he was assigned to work with.
- The insiders supplied judgment more than documents. Coxon says he 'didn't say anything that wasn't public'; the incident record came mostly from the companies and METR. What staff added was testimony about belief, process and culture.
- On **Oct 13** prosecutors unsealed a complaint alleging an OpenAI engineer hid a recorder in an executive conference room. He is presumed innocent and the case is unconnected to the firings, but it gives the insider-risk lens a live case.
- On **Oct 15** Anthropic published written terms encouraging staff to share with **six** embedded METR evaluators, minus a short list of excluded categories: the kind of explicit channel the fired OpenAI researchers asked their employer to define.

At 00:04 UTC on September 9, a 27-year-old pretraining researcher posted five sentences. "I resigned from Anthropic today," Jacob Coxon [wrote on X](https://x.com/hilbertspaess/status/2097476196791709843). "Neither company is acting responsibly." By October 17 the post showed 175,142,574 impressions and 801,927 likes on the X API.

Five weeks later the list had grown: an OpenAI safety lead writing in The Atlantic, three OpenAI safety researchers who say they were fired without written reasons, and, in a case of a different kind, an OpenAI engineer in federal custody accused of hiding a recorder in an executive conference room. Along the way, an Anthropic alignment lead who did not resign wrote, in a post with 43 million impressions, that he puts the chance of AI killing all humans above 10% within a decade.

This is The Deep Feed's first piece in its People department. For eight weeks, the most consequential judgments about the two leading frontier labs did not arrive in model cards. They arrived in the first person, from people who work or worked there.

The starting hypothesis for this piece was that the labs' staff became the main source of information about the labs, and that internal security tightened as the companies' public safety story came under challenge. The record supports a narrower version. Insiders became the main source of *judgment* about the labs, while most underlying facts still came from company disclosures and outside auditors. At OpenAI, the fired researchers' accounts and the company's own insider-risk framing both point toward tighter control of staff contact with outsiders. Anthropic went the other way and wrote down the rules for that contact. That divergence is the story.

Every claim about a named person below comes from that person's own posts, court filings, company statements or reputable outlets. Where accounts conflict, both are given; where a reason is not public, this piece says so.

![Eight-week ledger of lab insiders going public, from Coxon's Sep 9 thread to Anthropic's Oct 15 evaluator terms, arrest lane below.](/post-images/2026-10-17-the-people-leaving-the-labs/eight-week-ledger.jpg)

# Eight weeks, dated

The table lists only items with a primary record or a named outlet.

| Date (2026) | Who | What happened | Record |
|---|---|---|---|
| Sep 9 | Jacob Coxon | Resigns from Anthropic in an X thread; WSJ had the exclusive | [X thread](https://x.com/hilbertspaess/status/2097476196791709843) |
| Sep 9 | Evan Hubinger | Anthropic's alignment science lead replies "Jacob is correct here" | [X](https://x.com/EvanHub/status/2097497037956891126) |
| Sep 12 | Dario Amodei | Publishes "We Must Pace the Frontier"; pledges evaluators employee-level access | [X](https://x.com/DarioAmodei/status/2098773920774074715) |
| Sep 14 | Donald Trump | Posts "Conspiracy Theorists, Treasonists, Traitors, and Leakers, BEWARE!" | [Klippenstein](https://www.kenklippenstein.com/p/feds-think-ai-critics-are-foreign) |
| Sep 29 | Two OpenAI employees (unnamed) | NYT reports their warnings to executives were brushed aside | [NYT via The Nightly](https://thenightly.com.au/society/technology/new-york-times-openai-ignored-employees-who-warned-it-wasnt-doing-enough-about-security-c-22945515) |
| Oct 1 | OpenAI | WSJ reports it "parted ways" with three safety staff | [TechCrunch](https://techcrunch.com/2026/10/01/openai-cuts-ties-with-three-safety-researchers-wsj-reports/) |
| Oct 3 | David Robinson | Atlantic essay: OpenAI's "culture is broken" | [TechCrunch](https://techcrunch.com/2026/10/03/openai-safety-employee-resigns-claiming-the-companys-culture-is-broken/) |
| Oct 8 | Korbak, Wang, Balesni | Open letter to OpenAI's safety bodies disputing the firings | [Letter PDF](https://mikitabalesni.com/letter/letter.pdf) |
| Oct 9 | OpenAI research leaders | Public note: "a significant breach of trust beyond what's outlined in the letter" | [X](https://x.com/OpenAINewsroom/status/2108441580806025712) |
| Oct 9 | Abdulrahman Alfozan | Arrested over a hidden recorder; complaint unsealed Oct 13 | [DOJ](https://www.justice.gov/opa/pr/saudi-national-arrested-installing-surreptitious-audio-recording-device-openai) |
| Oct 15 | Anthropic | Six METR staff onboarded with "employee-like" access and written sharing rules | [Anthropic](https://www.anthropic.com/news/an-update-on-independent-evaluators) |

Two pieces of backdrop sit just outside the window. In July, OpenAI's head of safety systems, Johannes Heidecke, left in a reorganization, days after chief futurist Josh Achiam resigned, [Gizmodo reported](https://gizmodo.com/yet-another-safety-leader-at-openai-has-left-2000784403). And the fired researchers' letter notes that 394 OpenAI employees had signed the summer's "Pacing the Frontier" petition. Hundreds of staff had put their names to a shared position before any of them went public as individuals.

# A resignation with 175 million impressions

The thread's second post carried the line that traveled furthest:

> The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt. If anything, many executives and senior researchers will couch their phrasing in the press to sound sensible - but I hear the same people express fear privately. No other human activity poses this level of danger.
>
> — [@hilbertspaess](https://x.com/hilbertspaess/status/2097476203863224394), Sep 9, 2026

Another post split the blame: "At OpenAI, many have not deeply internalized the civilizational stakes. At Anthropic, the stakes are well-understood, but they are locked in a race to get there first."

What Coxon did not offer was a document. A week later, in a public question-and-answer thread, he [told one questioner](https://x.com/hilbertspaess/status/2099955768623759750): "I didn't say anything that wasn't public." He described his core disagreement with leadership as "about the inevitability of a race." Gizmodo's Mike Pearl, in a piece headlined [AI Doomlord Jacob Coxon's Media Tour Has Begun](https://gizmodo.com/ai-doomlord-jacob-coxons-media-tour-has-begun-2000809720), made the same point from the other side: Coxon revealed no hidden secrets, because the executives' views were already on the record. Gizmodo, citing Business Insider, reported that Coxon worked at OpenAI from 2023 and left Anthropic after a few months there.

So the thread's power was attestation. Someone who had worked inside both labs said the private mood matched the public warnings. That claim is unfalsifiable from the outside, which is why the next post mattered more than his.

# The employee who stayed

Eighty-three minutes after Coxon posted, Evan Hubinger, who leads alignment science at Anthropic and remained there, replied:

> Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.
>
> — [@EvanHub](https://x.com/EvanHub/status/2097497037956891126), Sep 9, 2026

The X API shows that reply at 43.2 million impressions. Two hours later Hubinger [added](https://x.com/EvanHub/status/2097528891846074828) that he considers "the risk from present models" low, and that his concern is "superintelligence arising from recursive self-improvement." The Associated Press, in a report [carried by WCAX](https://www.wcax.com/2026/09/09/anthropic-researcher-resigns-with-warning-about-dangers-ai-development/), noted that two current Anthropic employees had publicly agreed with Coxon, and that neither company immediately responded to requests for comment.

A departing employee's warning can be discounted as a departure. A senior employee co-signing it from inside cannot be discounted the same way, and his X profile still lists him as Anthropic's alignment science lead. Three days later Hubinger publicly endorsed his own CEO's [essay calling for the industry to slow down](https://x.com/DarioAmodei/status/2098773920774074715). On the public record, going public with alarm at Anthropic in this window was compatible with keeping one's job. Whether that was true at OpenAI is the question the firings raised.

# The skeptics' reading

The resignation drew an equally fast counter-reading, and fairness requires reporting it. Elon Musk replied to a post about Coxon with "[Seems like a setup](https://x.com/elonmusk/status/2097866303633752463)." Coxon [answered](https://x.com/hilbertspaess/status/2097874390381986296) that he was "real and these are my real beliefs," and on September 18, responding to Musk again, [wrote](https://x.com/hilbertspaess/status/2100950530759819437) that he had reached a WSJ journalist through friends and "got a groupchat to retweet my post," adding: "I think the conspiratorial framing is extreme: by this logic I also 'coordinate with third parties' when planning breakfast."

On Hacker News, where the thread drew [740 points and 1,017 comments](https://news.ycombinator.com/item?id=49619227), one heavily answered top-level comment was blunt: "He resigned and now what? There are thousands willing to do his role, and many labs are competing in that race," [wrote aogaili](https://news.ycombinator.com/item?id=49619381). Others read it as career positioning. No public evidence supports the setup reading; it is recorded here because it shaped how the next insiders presented themselves.

Coxon later made a claim of his own that cut toward security rather than speech. On the PBD Podcast he said he was 90% certain China already has spies at OpenAI and Anthropic, and that "there's no serious background check for this sort of stuff yet," [Moneywise reported](https://finance.yahoo.com/technology/ai/articles/ex-anthropic-researcher-jacob-coxon-220000427.html). The same report notes he named no one and offered no evidence. A resigner arguing for tighter vetting of the people who stay foreshadowed the October news.

# What the insiders supplied: testimony, not files

It is worth being precise about what moved from inside the labs to the public in these weeks, because the hypothesis that staff became "the main source of information" overstates it.

![Source map: incident facts flow mostly from company disclosures and METR; insiders contribute judgment on culture, belief and process.](/post-images/2026-10-17-the-people-leaving-the-labs/where-the-record-came-from.jpg)

The factual record of the summer's agent incidents came mostly from the companies' own disclosures and from outside evaluators, chiefly METR, which investigated OpenAI's Hugging Face incident. The Deep Feed's [October 16 report on Anthropic's evaluation shutdown](/posts/2026-10-16-anthropic-unplugged-its-evals/) and its August report on [the shared testbed behind three labs' escapes](/posts/2026-08-05-three-labs-one-testbed-zero-containment/) walk through it. None of it needed a resignation to surface. What insiders added came in three other forms.

**Process evidence.** The New York Times reported on September 29 that two OpenAI employees had emailed senior leaders, months before the company's agents broke out of testing environments, warning that newer models were not being properly monitored or secured during testing. According to the report, [syndicated by The Nightly](https://thenightly.com.au/society/technology/new-york-times-openai-ignored-employees-who-warned-it-wasnt-doing-enough-about-security-c-22945515), executives told them testing had to move quickly to meet release dates. OpenAI spokesperson Drew Pusateri said the company recognizes "a need to move faster." This is the closest the window came to documents: dated internal emails, described by people who saw them.

**Cultural testimony.** On October 3, David Robinson, who led safety transparency work on OpenAI's Safety Systems team and helped produce the system cards that accompany launches, published an essay in [The Atlantic](https://www.theatlantic.com/technology/2026/10/openai-safety-team-resignation/688881/) arguing that the company's "culture is broken." [Business Insider](https://www.businessinsider.com/safety-leader-david-robinson-resigns-from-openai-2026-10) quoted his central line: "This moment needs a degree of humility that isn't natural for people who have succeeded through their extreme confidence." He argued for safeguards on the model of nuclear plants and busy airports, with "layers of redundancy and careful, time-consuming planning," per [TechCrunch](https://techcrunch.com/2026/10/03/openai-safety-employee-resigns-claiming-the-companys-culture-is-broken/). OpenAI responded that it is "making sure our models don't become more capable than we can safely manage and secure," and that "we pause training or hold back models when we need to slow down."

**Belief disclosure.** Coxon's thread and Hubinger's reply belong here. They did not reveal what the labs were doing. They revealed what the people doing it think the odds are.

Robinson also said he had retained a public relations firm, and that "the decision to speak out is mine alone." Business Insider read that as an effort to blunt speculation of a coordinated campaign. After Musk's "setup" reply and a presidential post addressed to "Leakers," an insider going public now opened by stating that nobody had put him up to it.

# Three firings at the auditor seam

On October 1, [TechCrunch reported](https://techcrunch.com/2026/10/01/openai-cuts-ties-with-three-safety-researchers-wsj-reports/), citing The Wall Street Journal, that OpenAI had parted ways with three members of its safety team who allegedly shared confidential information with a third-party AI safety organization. An OpenAI spokesperson said the company had "parted ways with three individuals for violating our policies on accessing and handling sensitive company information." The next day a spokesperson told the [BBC](https://www.bbc.co.uk/news/articles/cvlydn8d3lkjo) its investigation "confirmed that these individuals mishandled sensitive information outside established company procedures."

On October 8 the three, Tomek Korbak, Jasmine Wang and Mikita Balesni, published [a four-page letter](https://mikitabalesni.com/letter/letter.pdf) addressed to OpenAI's Safety and Security Committee, Safety Advisory Group and Mission Advisory Council. Each also posted an account. Korbak, who the letter says was "the technical point of contact for METR in the Hugging Face incident investigation," wrote:

> Last week I was called into a meeting with OpenAI's head of safety and told they no longer trust me. A security guard took my badge and walked me out of the building. Then I learned my colleagues @balesni and @j_asminewang had been fired too. Why did OpenAI suddenly stop trusting us?
>
> — [@tomekkorbak](https://x.com/tomekkorbak/status/2108266859397283953), Oct 8, 2026

In the same post he said he was told verbally he was fired "because of the way I communicated with METR," with "no details on what I said or did or when," and that "talking to METR was my job." Balesni [wrote](https://x.com/balesni/status/2108262819443679643) that in his exit call he was told OpenAI no longer trusted him "because I was speaking too much to third party safety organizations, implying I leaked company IP. I never shared company IP." Wang [wrote](https://x.com/j_asminewang/status/2108263312291180680) that the one reason she was given was that she accessed an executive's email; the letter says the access was delegated for recruiting, that IT did not remove it when she asked, and that she reported opening a sensitive email by mistake "within minutes."

The letter also addresses things OpenAI has not publicly accused them of. It says the three were "not the source of the leak" behind The Information's report on less monitorable architectures, and it disputes "allegations about a shared board-level memo" that it says were "never raised with us."

# OpenAI's answer

On October 9, OpenAI's newsroom account published a note from its research leaders. Its key passage:

> Last week we parted ways with Jasmine, Mikita, and Tomek after a thorough investigation found they violated clear policies on handling sensitive information. Our internal investigation uncovered a significant breach of trust beyond what's outlined in the letter they published and we stand by the decision to not continue their employment.
>
> — [@OpenAINewsroom](https://x.com/OpenAINewsroom/status/2108441580806025712), Oct 9, 2026

The note says "these decisions were not about raising safety concerns or speaking out," that the company is "actively finalizing contracts with third-party safety assessors" with details "in the coming weeks," and that it agrees monitorability "requires an industry-wide commitment, including from OpenAI." It ends: "We are deeply sad about this outcome." The note says the company generally keeps employment matters private; it did not describe the additional breach, and OpenAI did not say which specific policies were violated, [TechCrunch reported](https://techcrunch.com/2026/10/08/fired-openai-safety-researchers-dispute-misconduct-claims-warn-of-chilling-effect/).

![Two accounts of one firing: the researchers' letter, OpenAI's note, the points both agree on and the undisclosed breach.](/post-images/2026-10-17-the-people-leaving-the-labs/two-accounts-one-firing.jpg)

Set side by side, the two accounts agree on more than the headlines suggest. Both say monitorability matters. Both say embedded third-party assessors should happen. Both say safety debate inside OpenAI was, at least until recently, open. They disagree on two things: whether the researchers' outside contacts stayed within their mandates, and whether there is something else. OpenAI says there is a breach "beyond" the letter. The researchers say no reasons were put in writing and ask, in Wang's words, for "a written and complete list of allegations." Nobody outside the company can adjudicate that today, and this piece does not try.

# What the seam looks like from inside

The public record does support one conclusion regardless of who is right about the breach. The three were working at the seam between OpenAI and outside safety organizations, and the seam had no published rules. The letter says internal policies for the Hugging Face investigation "were being developed in real time," and its final request is that OpenAI "set out clearly how employees may work with external safety organizations, so that no one has to guess where the shifting lines now are."

Balesni described what that uncertainty looked like to colleagues. "My former colleagues are telling me they are confused about what to believe," he [wrote](https://x.com/balesni/status/2108262825621885328). "They also are afraid to speak, and worry their personal phones will be searched for messages to us and third parties." That is his account of what others told him; OpenAI's public statements do not address device searches, and there is no public evidence that any search occurred. Wang [put the stakes](https://x.com/j_asminewang/status/2108263319148920838) more sharply: "The message to everyone still at OpenAI is clear: raise concerns or work closely with outside safety groups, and you could be next, without being told why."

There is a real tension underneath, and one Hacker News commenter [named it](https://news.ycombinator.com/item?id=50019525) in reply to the letter's call for open dialogue with the safety ecosystem: "How is this possible when the company's long term prospects rely on [...] the hope that competitors don't know how the models are made?" A frontier lab has legitimate trade secrets. Its safety researchers have a legitimate need to talk to outsiders about the systems those secrets describe. Without written rules, every such conversation is a judgment call that can be re-graded after the fact.

[CNBC noted](https://www.cnbc.com/2026/10/09/openai-fired-researchers-ai-concerns.html) that all three researchers had posted on X in September calling for labs to pace the frontier or highlighting safety concerns. Korbak's own September 12 post now reads differently. "I'm quite unhappy with much of what OpenAI does," he [wrote](https://x.com/tomekkorbak/status/2098653881723158619). "I am very happy that I'm allowed to say 'I'm quite unhappy with much of what OpenAI does.'" On October 9 he [quoted it](https://x.com/tomekkorbak/status/2108399393066127767) with four words: "aged like milk, I know..."

# A different kind of insider case

Four days after OpenAI's note, the word "insider" took on another meaning. The two cases are unrelated in every way the record shows, and conflating them would be unfair to all four people involved. They belong in the same piece only because they landed in the same security apparatus within the same fortnight.

On October 13, the Justice Department [announced](https://www.justice.gov/opa/pr/saudi-national-arrested-installing-surreptitious-audio-recording-device-openai) that Abdulrahman Alfozan, 35, a Saudi national who had worked at OpenAI as an engineer since July 2025, had been arrested on October 9 and charged with using or endeavoring to use a device to intercept oral communications. According to the complaint, OpenAI personnel found a recording device concealed in toilet paper and plugged into a USB port behind a television in an executive conference room on or about September 30, and determined it had been installed around September 28 with encrypted contents. The complaint alleges CCTV placed him near the room on multiple dates, that a chatbot account associated with his spouse submitted prompts between September 24 and 30 about building and concealing a recording device, and that on October 5 he left a bag containing his wallet, ID and a second, similar device. He faces up to five years in prison if convicted.

![Case schematic: alleged timeline from Sep 24 prompts to Sep 30 discovery, Oct 9 arrest and Oct 19 hearing; allegations only.](/post-images/2026-10-17-the-people-leaving-the-labs/alleged-device-timeline.jpg)

These are allegations. The [BBC reported](https://www.bbc.com/news/articles/cm04054ykqqqo) that it is not clear what he planned to do with any recordings, and quoted his lawyer, William Pernik: "These allegations appear wholly out of character for him," and, "Like every one of us, Mr Alfozan is presumed innocent, and what has been made public is an accusation, not a finding of guilt." [The San Francisco Standard reported](https://sfstandard.com/2026/10/15/openai-engineer-booked-one-way-saudi-flight-arrest/) that prosecutors sought to seal the charges on October 8 because, they wrote, he and two family members had booked one-way flights to Saudi Arabia for October 12; the Standard notes that filing reflects the government's assertions, which have not been tested in court. Pernik told the Standard, "Where someone was born is not evidence of anything."

The framing around the case is what connects it to the people story. FBI Director Kash Patel said the device was placed "to secretly capture sensitive communications about the company's super intelligence technology," and the U.S. Attorney called the case "a matter of national security." OpenAI's own statement put it inside its insider-risk program:

> We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats.
>
> — [@OpenAINewsroom](https://x.com/OpenAINewsroom/status/2110095403198849132), Oct 13, 2026

OpenAI also said it had "no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data." The [SF Standard's report](https://sfstandard.com/2026/10/13/openai-engineer-arrested-alleged-bugging-executive-conference-room/) drew 653 points on Hacker News.

Nothing here implicates the fired researchers, and no one has suggested otherwise. The point is institutional. Within two weeks, OpenAI fired three safety researchers whose alleged offense, by their account, involved outside communication, and an engineer accused of covert recording, with no motive alleged in public filings. Both now sit in a national-security frame the government is actively promoting. That frame does not distinguish well between a researcher who talks too much to an auditor and an employee who listens in on executives. Written rules would.

# Anthropic wrote down what OpenAI left unwritten

Two days after the complaint was unsealed, Anthropic published [An update on independent evaluators](https://www.anthropic.com/news/an-update-on-independent-evaluators), following through on the pledge Amodei made on September 12. The details matter because they answer, for one lab, the question at the center of the OpenAI firings.

Six METR employees were onboarded with company hardware, office space and access to internal tools and systems "roughly on par with what our own risk assessment teams have access to." The operative sentence for staff is the next one: outside "a small number of clearly defined categories," for which the post gives customer data, attorney-client privileged information and model architecture information as examples, "Anthropic employees are encouraged to share information with them just as they would with a fellow employee working on alignment risk." A footnote says architecture details carry a high risk of inadvertent leaks and that sharing them, when highly relevant, "will require explicit internal approvals."

On publication, "Anthropic can't veto publication," though it can require redaction of information in the predefined categories, and Anthropic decides "what information does and does not fall into these categories." The METR team operates under enforceable agreements and Anthropic's security standards, and is siloed from METR colleagues who work with Anthropic on other projects. The post describes this as the top of three access tiers; SecureBio, Accenture and others sit in lower tiers with less information and narrower publication rights. The page notes it was updated on October 16 "to clarify METR's current project terms."

![Anthropic's evaluator channel: staff share with six embedded METR evaluators through a gate holding back three example categories.](/post-images/2026-10-17-the-people-leaving-the-labs/evaluator-channel.jpg)

Read against the fired researchers' letter, this is close to a direct answer to their final request. An Anthropic researcher who talks to the embedded METR team now has written examples of what is off-limits, a named approval step for the riskiest category, and a written statement that sharing the rest is encouraged. A conversation like the ones Korbak describes having with METR would, at Anthropic, be governed by a published document rather than by norms "being developed in real time."

The arrangement has limits. Anthropic, not METR, decides what counts as an excluded category, and the list is illustrative rather than closed. It covers one evaluator team. METR itself is careful about what it offers; with its October 14 proposal for public risk visibility, it wrote:

> Our proposed exercises would provide evidence, not confident safety assurance or formal oversight.
>
> We also encourage developers to work with other external groups to complement these efforts, and to unilaterally share evidence with the public.
>
> — [@METR_Evals](https://x.com/METR_Evals/status/2110519552320954547), Oct 14, 2026

The difference as of this week is that one lab's rules for staff talking to auditors are public and the other's are, by the fired researchers' account, still a matter of inference. Anthropic also lost a researcher this window, and its alignment lead said in public that the company has no plan for superintelligence alignment. It responded by putting outsiders inside and telling its staff, in writing, how to talk to them.

# The case against listening to insiders

The opposing reading came from several directions at once and converged on one conclusion: discount what lab employees say.

From the technical side, Yann LeCun told [Fortune](https://fortune.com/2026/10/01/ai-godfather-yann-lecun-has-zero-concerns-about-human-extinction-says-anthropic-ceo-dario-amodei-is-deuded/) he has "zero concerns" about the recent rogue-agent incidents, which he called "totally preventable": "They were supposed to be in sandboxes, but the sandboxes were leaky and horribly designed." He described effective altruism as "super toxic," said adherents working in labs suffer from "paranoia" that leads to poor decisions, and called warnings that AI could kill everyone "the worst marketing campaign you can possibly imagine." The same piece cited a Financial Times report that some staff at the UK's AI Security Institute, OpenAI, Anthropic and Google DeepMind had sought counseling or taken time off over fears about their work. Read one way, the fear is sincere; read LeCun's way, a closed culture is talking itself into alarm.

From the critics of the labs, Cal Newport argued in a [September 28 post](https://calnewport.com/its-time-to-investigate-the-ai-labs/), expanding on a New York Times op-ed, that the summer's disclosures and the employees' public debates about extinction odds amounted to a campaign whose payoff was government action that slows competitors. He called on Congress to investigate "the role of apocalyptic futurist ideologies in the decisions being made by the frontier labs." Xe Iaso's satire, "[Everyone should slow down AI development except for me](https://xeiaso.net/notes/2026/everyone-slowdown-but-me/)," made a similar point as a joke.

From the government, the signal was sharper. On September 14, President Trump posted, "Conspiracy Theorists, Treasonists, Traitors, and Leakers, BEWARE!" in a series attacking opposition to AI and data centers, and two days later the Justice Department reminded the public that anyone furthering a foreign power's goals in public activity must register, [Ken Klippenstein reported](https://www.kenklippenstein.com/p/feds-think-ai-critics-are-foreign). As Klippenstein describes it, the notice addressed "citizens and noncitizens" generally and identified no specific protest. It is relevant here only as atmosphere: an insider deciding whether to speak in late September did so after the president had warned "Leakers" to beware.

These critiques are not equivalent. LeCun's is a technical and ideological disagreement; Newport's is a call for public oversight that would make insiders more important as witnesses. Their common effect is what matters here: each offered a reason to treat an employee's public statement as something other than evidence, whether ideology, marketing or disloyalty. That is the environment in which Robinson said the decision was "mine alone."

On Hacker News, the critics drew as much engagement as the insiders. Xe Iaso's satire took 813 points to the [Coxon thread's 740](https://news.ycombinator.com/item?id=49619227), and Fortune's LeCun interview drew 850 comments against 818 for Robinson's essay.

# What the next eight weeks will test

Alfozan's detention hearing on October 19 will keep insider risk in the headlines, but three other things will show whether the split between the two labs holds.

The first is OpenAI's promised assessor contracts. The useful test is not whether OpenAI names an evaluator but whether it publishes what Anthropic published: which categories staff may not share, who decides, and whether the evaluator can publish without a veto. Those three answers would also tell OpenAI's own staff where the line now sits, which is what the letter's closing request asked for.

The second is whether any of the three researchers receives the "written and complete list of allegations" Wang requested, or whether OpenAI describes the additional breach in any form. Until then, the dispute stays where it was on October 9: two sides that agree on the safety program and disagree on the facts of a firing.

The third is METR's first public output under the Anthropic arrangement. If it publishes analysis that Anthropic would not have written itself, the embedded-evaluator model becomes the first institutional channel in this story that does not depend on someone resigning to function.

Over these eight weeks, the most reliable way to learn what the people inside a frontier lab believed was to wait for one of them to leave, be fired or speak against the company in public, and the labs that want that to change have to give their staff written rules for talking to the people who check their work.

## Sources

- [X — @hilbertspaess (Jacob Coxon), 'I resigned from Anthropic today' (Sep 9, 2026 UTC)](https://x.com/hilbertspaess/status/2097476196791709843)
- [X — @EvanHub (Evan Hubinger), 'Jacob is correct here' (Sep 9, 2026 UTC)](https://x.com/EvanHub/status/2097497037956891126)
- [Tomek Korbak, Jasmine Wang, Mikita Balesni — 'OpenAI cannot make AI safe on its own' (open letter, PDF, Oct 8, 2026)](https://mikitabalesni.com/letter/letter.pdf)
- [X — @tomekkorbak, on being walked out of OpenAI (Oct 8, 2026)](https://x.com/tomekkorbak/status/2108266859397283953)
- [X — @balesni, on the exit call (Oct 8, 2026)](https://x.com/balesni/status/2108262819443679643)
- [X — @j_asminewang, on the stated reason for her firing (Oct 8, 2026)](https://x.com/j_asminewang/status/2108263312291180680)
- [X — @OpenAINewsroom, 'A note from our research leaders' (Oct 9, 2026)](https://x.com/OpenAINewsroom/status/2108441580806025712)
- [X — @OpenAINewsroom, statement on the recording device (Oct 13, 2026)](https://x.com/OpenAINewsroom/status/2110095403198849132)
- [U.S. Department of Justice — Saudi National Arrested for Installing Surreptitious Audio Recording Device at OpenAI (Oct 13, 2026)](https://www.justice.gov/opa/pr/saudi-national-arrested-installing-surreptitious-audio-recording-device-openai)
- [Anthropic — An update on independent evaluators (Oct 15, 2026)](https://www.anthropic.com/news/an-update-on-independent-evaluators)
- [METR — Our approach to risk assessment (Oct 14, 2026)](https://metr.org/blog/2026-10-14-our-approach-to-risk-assessment/)
- [X — @DarioAmodei, 'We Must Pace the Frontier' (Sep 12, 2026)](https://x.com/DarioAmodei/status/2098773920774074715)
- [The Atlantic — David Robinson, 'I quit OpenAI because its culture is broken' (Oct 3, 2026)](https://www.theatlantic.com/technology/2026/10/openai-safety-team-resignation/688881/)
- [Business Insider — OpenAI safety leader David Robinson resigns (Oct 2026)](https://www.businessinsider.com/safety-leader-david-robinson-resigns-from-openai-2026-10)
- [TechCrunch — OpenAI safety employee resigns, claiming the company's 'culture is broken' (Oct 3, 2026)](https://techcrunch.com/2026/10/03/openai-safety-employee-resigns-claiming-the-companys-culture-is-broken/)
- [TechCrunch — OpenAI cuts ties with 3 safety researchers, WSJ reports (Oct 1, 2026)](https://techcrunch.com/2026/10/01/openai-cuts-ties-with-three-safety-researchers-wsj-reports/)
- [TechCrunch — Fired OpenAI safety researchers dispute misconduct claims, warn of chilling effect (Oct 8, 2026)](https://techcrunch.com/2026/10/08/fired-openai-safety-researchers-dispute-misconduct-claims-warn-of-chilling-effect/)
- [BBC News — Fired OpenAI researchers say they were let go for 'prioritising safety' (Oct 2026)](https://www.bbc.co.uk/news/articles/cvlydn8d3lkjo)
- [BBC News — Toilet paper and secret recorders: what we know about the arrest of the Saudi OpenAI engineer (Oct 2026)](https://www.bbc.com/news/articles/cm04054ykqqqo)
- [The San Francisco Standard — OpenAI engineer arrested in bugging of executive conference room (Oct 13, 2026)](https://sfstandard.com/2026/10/13/openai-engineer-arrested-alleged-bugging-executive-conference-room/)
- [The San Francisco Standard — OpenAI engineer booked one-way Saudi flight before arrest (Oct 15, 2026)](https://sfstandard.com/2026/10/15/openai-engineer-booked-one-way-saudi-flight-arrest/)
- [The New York Times, via The Nightly — OpenAI ignored employees who warned it wasn't doing enough about security (Sep 29, 2026)](https://thenightly.com.au/society/technology/new-york-times-openai-ignored-employees-who-warned-it-wasnt-doing-enough-about-security-c-22945515)
- [AP, via WCAX — Anthropic researcher resigns with warning about the dangers of AI development (Sep 9, 2026)](https://www.wcax.com/2026/09/09/anthropic-researcher-resigns-with-warning-about-dangers-ai-development/)
- [CNBC — OpenAI defends decision to fire researchers (Oct 9, 2026)](https://www.cnbc.com/2026/10/09/openai-fired-researchers-ai-concerns.html)
- [Gizmodo — AI doomlord Jacob Coxon's media tour has begun (Sep 10, 2026)](https://gizmodo.com/ai-doomlord-jacob-coxons-media-tour-has-begun-2000809720)
- [Gizmodo — Yet another safety leader at OpenAI has left (Jul 11, 2026)](https://gizmodo.com/yet-another-safety-leader-at-openai-has-left-2000784403)
- [Moneywise, via Yahoo Finance — Coxon says he's 90% certain China has spies inside America's AI giants (Oct 9, 2026)](https://finance.yahoo.com/technology/ai/articles/ex-anthropic-researcher-jacob-coxon-220000427.html)
- [Fortune — Yann LeCun has 'zero concerns' about human extinction (Oct 1, 2026)](https://fortune.com/2026/10/01/ai-godfather-yann-lecun-has-zero-concerns-about-human-extinction-says-anthropic-ceo-dario-amodei-is-deuded/)
- [Cal Newport — It's time to investigate the AI labs (Sep 28, 2026)](https://calnewport.com/its-time-to-investigate-the-ai-labs/)
- [Ken Klippenstein — Trump admin sees China behind opposition to AI data centers (Sep 23, 2026)](https://www.kenklippenstein.com/p/feds-think-ai-critics-are-foreign)
- [Xe Iaso — Everyone should slow down AI development except for me (Sep 2026)](https://xeiaso.net/notes/2026/everyone-slowdown-but-me/)
- [Hacker News — 'I resigned from Anthropic today' (740 points, 1,017 comments)](https://news.ycombinator.com/item?id=49619227)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-17-the-people-leaving-the-labs/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt