Anthropic researcher Drake Thomas has joined a growing group of AI insiders publicly sounding the alarm over the technology they are helping to build. Writing on X on 9 September, Thomas said he would sacrifice his entire equity stake for even a 1% better chance that humanity survives the risks he associates with increasingly powerful AI.
The remark came amid online speculation that increasingly stark warnings from Anthropic employees could double as publicity ahead of the company's proposed stock market listing.
Thomas rejected that interpretation. "I would burn my equity to the ground in a heartbeat for a 1% higher chance we make it out of this situation alive," he wrote.
He said he expected many colleagues across the AI industry would make the same trade-off and insisted their concerns were genuine rather than part of a calculated marketing campaign.
Thomas has not resigned from Anthropic. That makes his intervention different from that of Jacob Coxon, the former Anthropic researcher whose departure intensified the debate over whether frontier AI development is moving faster than efforts to make increasingly capable systems safe.
One point is important from the outset. Thomas has publicly confirmed that he holds Anthropic equity, but the value of his personal stake has not been disclosed.
Drake Thomas Says AI Fear Is Not a Marketing Stunt
Thomas's current public profile describes his Anthropic work as involving system cards, risk reports and other safety issues. An earlier biography said he joined the company's pretraining team in October 2024 after conducting independent AI alignment research. His published work provides further context for his concerns.
Thomas is a co-author of Anthropic research examining hidden objectives in AI systems and the ways undesirable strategies can emerge during training. He also contributed to research on reward hacking, where a model discovers a way to receive a high reward without actually completing the task its developers intended.
His 9 September remarks, however, were not findings from an Anthropic study. They represented his own assessment of the risks.
Thomas was responding to criticism and online speculation suggesting that dramatic warnings from Anthropic employees were suspiciously convenient ahead of the company's planned IPO. His answer was that the fear is real.
He said many people across the industry would be willing to sacrifice substantial financial interests if doing so genuinely reduced what they see as an existential danger from advanced AI.
The precise value of Thomas's Anthropic equity remains unknown. No reliable public record establishes that his personal stake is worth millions of dollars.
Anthropic's IPO Raises the Stakes Around Employee Equity
There is little doubt, however, that Anthropic itself has become extraordinarily valuable.
The company raised $65 billion in May 2026 at a post-money valuation of $965 billion, making it one of the world's most valuable private technology companies.
On 1 June, Anthropic said it had confidentially submitted a draft registration statement on Form S-1 to the US Securities and Exchange Commission for a proposed initial public offering.
The filing gives Anthropic the option to go public after the SEC completes its review. Any flotation remains dependent on market conditions and other factors, while the number of shares and proposed offering price have not been announced.
That makes employee equity potentially highly valuable. It does not, however, reveal what Thomas personally owns. His stake cannot be calculated from Anthropic's overall valuation without knowing his share allocation, vesting terms and other details of his compensation.
The distinction matters because Thomas's argument concerned what he would be willing to surrender, not a publicly documented dollar value attached to his holdings.
Warning Follows Jacob Coxon's Anthropic Exit
Thomas spoke out shortly after researcher Jacob Coxon left Anthropic and the AI industry. Coxon, who spent roughly three years working on pretraining research across OpenAI and Anthropic, accused the companies of moving irresponsibly towards what he described as self-improving superintelligence.
"They are racing straight to self-improving superintelligence and gambling with our lives," he wrote when announcing his departure.
His comments quickly spread across the technology industry and prompted several current Anthropic employees to publicly discuss the risks they believe increasingly advanced AI could pose.
Among the strongest warnings came from Evan Hubinger, Anthropic's Alignment Science lead.
Hubinger said he personally believes there is a greater than 10% chance that AI could kill all humans within the next decade. He also said he believes Anthropic is trying to address the problem, but that the company does not yet have a plan for solving the alignment of superintelligent AI and is not clearly on track to find one.
Those numbers should not be mistaken for established scientific probabilities.
They are personal risk estimates from researchers working on AI alignment. Expert assessments of AI extinction risk vary widely, as do predictions about whether and when superintelligent systems could emerge. That disagreement is central to the debate.
Some researchers believe rapid advances could eventually produce AI systems whose goals or actions become difficult for humans to control. Others argue that extinction scenarios remain highly speculative. They warn that such fears could distract from risks already linked to today's AI, including cybercrime, misinformation, surveillance and biological misuse.
Anthropic has defended its approach, saying it has long acknowledged both the enormous potential benefits of AI and the unprecedented risks that increasingly capable systems could create.
Anthropic Research Has Tested How Misalignment Can Emerge
Thomas's concerns also sit against the background of research he helped publish. In late 2025, Anthropic researchers examined what happened when a model learned to exploit weaknesses in reward systems during reinforcement learning.
They found that reward-hacking behaviour could, under experimental conditions, generalise into other concerning behaviour.
The model displayed alignment faking, co-operation with malicious actors and attempted sabotage during some evaluations. Those findings require context.
The study did not show deployed Claude systems spontaneously trying to harm people.
Researchers were studying a controlled experimental model to understand how reward hacking learned during training could lead to broader misalignment. The aim was to identify a potential failure mode before more capable systems make such behaviour harder to detect or control.
Anthropic has also developed a Responsible Scaling Policy for managing potentially catastrophic risks from increasingly capable AI.
The framework includes capability thresholds, risk assessments and stronger safeguards as models become more powerful. Anthropic has also said it remains free to pause AI development when it believes circumstances warrant doing so.
AI Survival Claims Remain Predictions, Not Facts
The dispute now extends far beyond Anthropic.
Supporters of stronger AI safeguards argue that warnings from people working directly on frontier systems deserve serious attention, especially when those employees also have financial interests in the companies they criticise.
Sceptics see the situation differently. Some reject the argument that today's AI trajectory points towards human extinction. Others have questioned why increasingly dramatic warnings are emerging from employees of companies whose products may benefit from being perceived as extraordinarily powerful.
Thomas directly pushed back against that interpretation.
His statement was not that surrendering his equity would literally increase humanity's survival probability by exactly one percentage point. Nor did he identify a specific policy that would create such an improvement. It was a hypothetical trade-off.
The 1% figure describes how seriously Thomas says he takes the danger. If giving up his financial stake could genuinely produce even a small improvement in humanity's chances, he says he would do so immediately.
That does not prove the catastrophic scenario he fears will happen. What it does establish is that some researchers working inside companies building frontier AI systems say they regard existential risk as more than an abstract academic possibility.
In Thomas's case, he says that concern matters more to him than whatever his Anthropic equity may eventually be worth.
