Anthropic's Former Researcher Testifies at the New York City Council: What He's Saying and Who's Listening

One phone call brings an extinction warning into the hearing room

Former Anthropic researcher Jacob Coxon quit last month and dropped the line, "By the end of this decade, AI could drive the entire human species to extinction." He is now sitting in front of lawmakers as a named individual for the first time.

Coxon left Anthropic in September 2026. On his way out, he delivered a line that has been quoted repeatedly: "People seriously working on AI are convinced that by the end of this decade, AI could cause the entire human species to go extinct." His remarks targeted both his former employer and OpenAI, with the phrasing "gambling with our lives."

The invitation came from New York City Council Speaker Julie Menin. She had previously called publicly for AI whistleblowers to drop their anonymity and step up to the microphone. Coxon is the first person to pick up that phone call.

The City Council's agenda goes beyond just listening. A package of AI bills is under review, and Coxon's testimony has been folded into this legislative debate. What exactly he says at the hearing—and whether he names names—remains an open question.

Federal legislation in line; New York City Council invites whistleblowers first

Related Senate bills have stalled repeatedly, and no unified federal legislation targets frontier AI. The New York City Council is now reviewing a package of AI bills.

There is no federal legislation specifically governing frontier AI(the class of models with the most powerful capabilities and the greatest potential risks). Related Senate bills have been introduced and shelved multiple times, with windows closing for years at a stretch. The pace at the local level is different: the New York City Council is reviewing a package of AI bills covering algorithmic discrimination, automated hiring decisions, facial recognition, and more, with shorter legislative cycles and a tighter hearing schedule.

According to Bloomberg, Council Speaker Julie Menin actively requested that AI whistleblowers appear and testify. Coxon accepted Menin's invitation, pinning a long-horizon judgment like "AI could drive humans to extinction" onto a piece of local legislation already in the pipeline: council members now have a named testimony entered into the hearing record, and the bill's "internal risk" narrative has a source.

When Congress can't move, the City Council turns hearings into a de facto entry point for frontier AI regulation: the legislative window, whistleblower testimony, and council members' attention get bundled into the same local bill. Menin is pulling in whistleblowers, not academics—academics talk principles; whistleblowers say, "I've seen it, I've said it, nobody listened."

He just reread the prospectus, out loud

Coxon sounds like he's dropping a sensational doomsday prediction, but the line was already written in Anthropic's own IPO prospectus. The twist is the shift in use: the same risk statement moves from a legal disclaimer to a named accusation.

After leaving Anthropic last month, Coxon said: "People seriously working on AI are convinced that by the end of this decade, AI could cause the entire human species to go extinct." Nearly the same wording appears in the IPO prospectus Anthropic filed with the SEC (U.S. Securities and Exchange Commission, the U.S. federal securities regulator), where the company lists "loss of control over AI leading to human extinction" as a standalone risk factor and also identifies the U.S. government's stance toward the company as an uncertainty that could affect customer relationships and ripple out to human civilization. Coxon accepted the invitation from New York City Council Speaker Julie Menin to testify at the hearing, where council members are reviewing a package of bills aimed at putting safeguards around the technology.

A prospectus is a disclaimer for investors—the heavier it reads, the more it looks like fulfilling a disclosure duty. A hearing is different: the speaker is a named former employee, the audience is council members and voters, and the consequence is slapping a "gambling with our lives" label on the company. A statement originally written to protect the company has been flipped into an accusation.

Anthropic positions itself on its website as a provider of "frontier AI research and products that put safety first," claiming to operate as a public benefit corporation with the goal of "securing AI's benefits and reducing its risks." The prospectus and the self-positioning exist side by side: promising safety on one hand, while writing "extinction" into the risk section on the other.

This move does more damage to the company than the original sentence ever could. A risk buried in a prospectus is abstract, future-facing, and hedged with "uncertainties." An accusation delivered from the witness table is concrete, directed at a company by name, and pinned on "you." Investors can stomach "possible human extinction" in an SEC filing. When a former employee reads the same line to the City Council, the questions the company faces shift: you admit the risk and keep accelerating—who picks up the tab?

Coxon dragged OpenAI in alongside Anthropic because both companies' prospectuses contain similar extinction-level risk language. An industry fear, self-disclosed internally, has now been publicly converted into a joint accusation against the industry itself.

The prospectus signed the words; the hearing demands a public reckoning

Coxon isn't saying anything Anthropic didn't already write down. What's new is that an internal risk disclosure is now a public showdown—and the hearing record is where it lands.

Watch three things. Whether Coxon, in his testimony, pins "extinction risk" to specific product names, training data, model evaluation processes, or release schedules—whether it's general rhetoric or details that can trigger legislative liability—will determine how damaging this round of testimony turns out to be.

Whether New York City's AI bill package adds whistleblower protection clauses, referencing minimum standards from federal whistleblower laws(laws that protect internal corporate whistleblowers from retaliation)—this directly determines whether the next Coxon will dare come forward. Whether Anthropic and OpenAI choose silence or a public response—both signed the words in their prospectuses, and whether they respond or not, it all goes into the hearing record.

Council Speaker Julie Menin actively invited whistleblowers to testify, and her intent is clear: she doesn't want industry self-regulation talking points—she wants testimony that can be entered into the legislative record. The package of AI bills under review at the New York City Council is trying to answer a simple question: who watches over the risks these companies acknowledge in their own prospectuses? Are the prospectus disclaimers enough? Coxon's testimony lands an outside piece of corroborating evidence on legislators' desks at a pivotal moment in the bill debate.

Anthropic wrote in its prospectus that "the U.S. government's stance will affect the company's relationship with customers." Coxon showing up at the New York City Council rather than Congress neatly adds an exception to that line: a regulatory vacuum at the federal level doesn't mean local governments sit still. City-level hearing conclusions can become templates for other states and cities, and can feed federal legislation when Congress stalls. Coxon's choice to speak at the city level, and the timing, were not random.

Signals to watch: if Coxon names a specific model or release decision in his testimony, and council members cite it in subsequent bill amendments, this round of testimony shifts from a media event to legislative evidence. If whistleblower protection clauses are added to the bill text, the hearing has produced real institutional output. Conversely, if the testimony stays at the abstract level of "extinction risk" and the bill amounts to general statements with no whistleblower provisions, this hearing is closer to a public-opinion test than a regulatory turning point.

No livestream, no statement — so what do you actually watch?

Coxon's allegations and Speaker Menin's bills are both in motion. The hearing itself has no public gallery and no stream, so the useful move is to fix the variables now and check them against coverage as it lands.

Coxon has not posted a prepared statement on any public channel. The room stays closed to outsiders. What's trackable is everything that surfaces afterward.

One: does he name products? So far he's used sweeping phrases like "gambling with our lives" — no specific models, features, or decisions. Anthropic's prospectus already self-discloses that "loss of control over AI could cause human extinction." Coxon has carried the company's internal risk language to the council table. The open question is whether he follows it with names: which product line, which deployment, which safety mechanism was skipped. That's the line between an insider revelation as a PR skirmish and one as an actionable accusation.

Two: do the bills add whistleblower protection? Speaker Menin pulling whistleblowers toward the stage is a political signal on its own. It only becomes a rule-setting hearing if the package under review includes anti-retaliation provisions, identity protections, and legal safe harbors for internal whistleblowers. Watch the bill texts on the council member list for newly added clauses.

Three: do Anthropic and OpenAI answer? Both have stayed silent on Coxon's allegations. Anthropic's prospectus tone is "we take this seriously." OpenAI already has a record of public questioning after its CEO was summoned in Australia. Silence is information too — when they speak, what they say, and to whom, will show how much tolerance they have for insiders testifying in public.