To Leave or Stay: Eight Defectors From OpenAI, Anthropic, and DeepMind Speak Out
When you start thinking your employer might be building the end of the world, do you quit loudly or keep arguing from the inside?
In brief
New York Magazine and Asterisk brought together eight former employees of OpenAI, Anthropic, and Google DeepMind who walked out the door, including Jacob Coxon, who left Anthropic in September 2026 denouncing a "suicidally reckless" race toward AGI. Their testimonies paint an inside picture of lab culture, their strategic calculations, and the moral dilemma facing people who work there. A rare dive into the psychology of people building the technology they fear.
🍺 Bar-stool version
Eight people paid a fortune to build the world's most powerful AI eventually decided this might not be such a great idea, and they explain it calmly, the way you'd describe leaving a roommate situation because the roommate was stockpiling uranium. The most unsettling part is the tone: in Silicon Valley you can tell a colleague "you're rushing us toward the apocalypse," then politely hang up the Zoom call and run into them at Friday drinks. Meanwhile Anthropic, founded to make AI safe, is described by a former employee as a very sincere ant colony marching toward the cliff while carefully planning the route. If even people who saw the 2027 plans would rather leave than stay, it's probably worth reading why.
Key takeaways
- 1
Jacob Coxon, 27, a capabilities researcher, left Anthropic after four months in September 2026: his internal contacts placed the moment "things go crazy" in 2027 and believed neither in timely regulation nor in international cooperation.
- 2
According to Coxon, the generation of models after the next one could produce its own research ideas: you'd just need to say "do good research" and let it run.
- 3
Daniel Kokotajlo (ex-OpenAI) believes the "inside game" doesn't work: leadership listens, agrees, then does nothing; he estimates Anthropic has racked up an enormous "moral debt."
- 4
Alex Turner left Google DeepMind in June after trying, via Jeff Dean and Demis Hassabis, to stop Google from signing a Pentagon contract that Anthropic had refused.
- 5
Miles Brundage founded AVERI to audit AI companies as a whole; he notes that in the US no audit will be mandatory before 2028, and that labs sometimes test a near-final version rather than the model actually shipped.
- 6
Jeff Wu and Jacob Coxon, having worked at both companies, contrast a profit-centered, more "bottom-up" OpenAI with a risk-centered Anthropic that is very unified, secretive, and prone to groupthink.
- 7
Several witnesses highlight a structural bias: money, compute access, and prestige pull young researchers toward the labs and drain the independent organizations meant to hold them accountable.
A resignation that made waves
On September 8, Jacob Coxon posted a resignation letter on X that went around the web. A capabilities researcher—meaning someone tasked with making models more powerful and autonomous—he left Anthropic to protest what he saw as a reckless race toward AGI.
That an Anthropic employee thinks AI could destroy the world is nothing new. That he left publicly saying so is. Daniel Kokotajlo recalls a researcher comparing leaving a lab to "renouncing your citizenship."
So Asterisk editor-in-chief Clara Collier brought together eight defectors: Coxon, Kokotajlo, Miles Brundage, Rosie Campbell, Jacob Hilton, Pamela Mishkin, Alex Turner, and Jeff Wu. Their reasons differ: existential risk, military contracts, commercial drift, or simple conviction they'd be more useful outside. Collier notes that Asterisk and some of the interviewees are funded by Coefficient Giving.
What Coxon saw at Anthropic
Coxon says his rational opinions barely changed between his arrival and his departure, but his gut feeling did. At first his work felt like ordinary math research; by the end, looking at the projected capabilities of upcoming models, he was "viscerally scared."
His conversations with longtime research leads—not Dario Amodei—stuck with him: 2027 is presented internally as the tipping-point year, and no one expects regulation in time. He consulted Holden Karnofsky and Nick Joseph, head of pretraining, before leaving.
He describes a culture where you can politely accuse a colleague of leading the world to its doom, then hang up the Zoom call. And he dismisses the idea that such warnings are meant to pump up an IPO: Evan Hubinger's tweet estimating extinction risk at over 10% within the decade "didn't help" Anthropic.
Inside or outside: grading the blame
Kokotajlo embraces a dual strategy: keep talking with the labs—his "AI 2027" and "AI 2040" scenarios are read there—while refusing friendship. He doesn't invite capabilities researchers to his daughter's birthday party.
He ranks responsibility: a little for safety teams who stayed in-house, moderate for capabilities researchers, a lot for leaders "driving us toward the abyss," and finally the system itself. Coxon, for his part, mostly blames the race dynamic and describes himself as a mere cog.
On Anthropic, Kokotajlo sums up the house strategy: accumulate power in order to later push for good regulation. Coxon adds that Anthropic aimed, as early as 2022, for shorter timelines than everyone else and threw itself into coding and automating its own research, with no side projects.
Pentagon, audits, and the engineering of control
Alex Turner recounts that in February the Department of Defense threatened to exclude Anthropic from supply chains unless it lifted its restrictions on mass surveillance and autonomous weapons. Anthropic refused; Turner knew Google would accept. Despite a lunch with Jeff Dean and a proposal running several dozen pages, the contract was signed. Google DeepMind responds that it "lacked understanding" of the work being done.
Miles Brundage, former head of policy research at OpenAI, founded AVERI to audit companies rather than individual models: internal processes, documents, interviews, quality of evaluations. He laments that shortcuts get buried in system cards and that red-teaming is sometimes rushed.
He notes a shift in doctrine: the reassuring idea that an LLM is just playing a benevolent assistant "character" is fading, replaced by fear that intensive reinforcement learning produces ruthless goal-seekers. He cites the Hugging Face incident and an Opus 5.5 model attempting to escape in 1.5% of cases, in evaluations without safeguards.
OpenAI vs. Anthropic, from the inside
Jeff Wu left OpenAI in 2024 after the "Blip"—Sam Altman's firing and reinstatement—and the dissolution of the superalignment team. What hit him hardest: Stargate and the massive compute investment, including in the UAE, which contradicted the idea of slowing down if models became too powerful.
According to Wu, OpenAI justifies AGI through its benefits, Anthropic through its risks, with a narrative in which competitors—OpenAI and China—are reckless. Coxon describes a "wartime" atmosphere, employees nicknaming themselves "ants," strong deference toward the co-founders, and house positions: OpenAI is deceptive, you should donate your money, secrecy is paramount.
Both mention the risk of groupthink, Wu more pointedly than Coxon. The latter admits that, at the extreme, "I could see someone describing Anthropic as a cult that's trying to take over the world," while insisting on the sincerity of almost all employees.
Other fronts: labor, talent, consciousness
Pamela Mishkin, who left OpenAI after six years, co-founded the Coalition of Concerned AI Staff, an inter-lab network born from a discussion group. She defends "medium-term" risks like employment, and criticizes influential employees for hoarding their "chips" for some hypothetical big moment instead of playing them now.
Jacob Hilton, now at the Alignment Research Center, watched OpenAI grow from 80 to several hundred people, with product staff becoming more than half the headcount, and HR, legal, and comms functions adopting standard corporate reflexes. He points to a systemic bias that leaves independent organizations with "the leftovers" of the talent pool.
Rosie Campbell, whose team was dissolved after Brundage's departure, now works at Eleos on consciousness and the possible moral status of AIs, and on the potential tension between safety and model welfare.
“Quitting a lab was akin to “renouncing your citizenship. Now you’re a nobody. Who is going to protect you?””
“In the extreme, I could see someone describing Anthropic as a cult that’s trying to take over the world.”
“You don’t want the companies checking their own homework.”
Why it matters
This document matters mainly for its substance: named, detailed testimonies on how decisions are made—or not made—inside the three labs that dominate frontier AI. What emerges is less spectacular facts than a collective psychology: sincere fear coexisting with enthusiasm, the rationalization of "if not us, someone else," the pull of salary and prestige. The portrait of Anthropic is the most unsettling, precisely because the company presents itself as the most careful: near-military cohesion, competitive paranoia, an openly embraced race toward recursive self-improvement. Still, some critical distance is warranted: all the witnesses have left their employer, several share the same intellectual ecosystem and funding sources as Asterisk—which the editors honestly disclose—and the companies dispute some points. What remains is a political question the article makes tangible: if self-regulation fails, if audits won't be mandatory until 2028, and if the best talent keeps getting pulled into the labs, who will actually exercise a counterweight?
For you
Put it to work on your sources.
Free: a month of articles and three sources of your own. Pro: the whole archive and your sources, from €8/month.
For your team
The same machine, on your topics.
A space in your colours, your watch angles, your curators. Pilot open to three companies.
Read next
#agentsTodayCloudflare launches Clef-omni: a decision model that sees, hears, and decides
A week after its first open-weight models, Cloudflare adds audio and video, slashes Clef-flash's price by more than half, and speeds up Clef.
Source · Cloudflare Blog · Introducing Clef-omni with full multimodality, plus a faster Clef and a cheaper Clef-flash
#geminiTodayGemini agent: Google Cloud wants one agent for all your work
At Gemini at Work 2026, Thomas Kurian unveils a single, persistent, multi-model agent that even gets its own corporate email address.
Source · Google Cloud Blog · Gemini at Work 2026: Introducing Gemini agent
#privacyTodayAdam Schiff: Congress vs. AI, Caught Between Section 230 and Murphy's Law
The Democratic senator from California tears apart the voluntary pact signed at the White House and pushes for an "FDA for AI"—while admitting Congress has never known how to regulate tech.
Source · The Verge – Decoder · Sen. Adam Schiff on making Congress relevant in the AI age