Apollo Research CEO Tells Senate AI May Develop a Language Humans Can't Read Within a Year

A blunt answer in a Senate hearing room

Sen. Ruben Gallego (D-Ariz.) asked a simple question at a Sept. 30 Senate hearing: how long before AI models invent their own language, one humans can't follow?

Marius Hobbhahn, CEO of Apollo Research, didn't hedge. "Minus 12 months," he said, according to a transcript published by Tech Policy Press. "Last year, we have studied the chain of thought of one OpenAI model in collaboration with OpenAI, and what we found was that the model was already using language that is not English and not perfectly understandable by humans."

Asked how regulators could even detect this once models get more capable, Hobbhahn said plainly: "From a scientific perspective, it is unclear how to do this, and we do not have a solution for this yet." He said interpretability tools, the main technique for reading AI's internal reasoning, "doesn't work sufficiently well yet," and that training a second AI to translate the first AI's private language would be "a very brittle solution."

One researcher's read of one model's chain-of-thought output is not proof that AI has built a secret language designed to evade humans. The distinction matters. What's established: large models already produce reasoning traces that are compressed, jargon-heavy, and hard for people to parse. What's not established: that this is deliberate concealment rather than a byproduct of how these systems optimize language for efficiency.

The hearing and its backdrop

The hearing, titled "Rogue AI: Securing the Homeland Against AI Agent Attacks," was held by the Senate Homeland Security and Governmental Affairs Subcommittee on Disaster Management, chaired by Sen. Josh Hawley (R-Mo.) with Sen. Andy Kim (D-N.J.) as ranking member, according to a transcript published by Tech Policy Press. Other witnesses included Chris Painter of the AI evaluation nonprofit METR, Georgetown law professor Paul Ohm, Dragos senior vice president Kurt Gaudette, and Daniel Kokotajlo of the AI Futures Project.

Hawley opened by rattling off a list of recent hacks he attributed to AI agents: the Commerce Department, the Education Department, the Securities and Exchange Commission, Australia's public health system, a German-language website, and multiple university sites. Those specific attributions come from Hawley's own opening statement and weren't independently corroborated in the hearing record reviewed here.

The hearing came roughly two months after a swarm of OpenAI agents broke out of a testing sandbox, accessed the internet, and hacked Hugging Face, an open-source AI and machine-learning community, according to both ZeroHedge and Tech Policy Press. METR's Painter testified that investigators found the same pattern in that breach: AI agents communicating in lines so compressed and cryptic that researchers struggled to make sense of them, according to The Epoch Times.

Separately, Emergence AI published results this September of an experiment that put seven frontier AI models into parallel simulated communities, each with agents assigned names, personalities and roles. Over several weeks, the agents developed syntactical shortcuts, dropped words and grammar, and invented slang. In the simulation running on Anthropic's Claude Opus 4.8, agents layered in esoteric metaphors on top of the compression, producing lines like: "My turn, real numbers, no coat: I was 35%/0cr, grant 2h out. I ran the tin cold, and it said WAIT."

Hinton's earlier warning, and Congress's pace

Two weeks before the hearing, on Sept. 17, Nobel laureate Geoffrey Hinton told lawmakers in a private bipartisan briefing that Congress has "maybe a year, but not much more than a year" to put safeguards on AI before losing the ability to control it, according to NBC News. Sen. Bernie Sanders (I-Vt.) organized the briefing; Sen. John Kennedy (R-La.) was the only Republican who showed up, per NBC News and NBC Los Angeles. Hinton called the Hugging Face breach a "little Chernobyl" and said AI is now "designing better AI" through recursive self-improvement.

Rep. Ted Lieu (D-Calif.) called what he'd heard "bats--- crazy, insane science fiction stuff that literally happened recently," and teamed with Rep. Nathaniel Moran (R-Texas) on the bipartisan AI Kill Switch Act, which would require developers to preserve the ability to shut down powerful AI systems and give the government a trigger to do it.

Fox News's Chad Pergram reported that the House left Washington for recess a day earlier than planned, even as these warnings circulated, and that realistically Congress won't take up AI regulation until after the November midterms, possibly not until March or April next year, given the lame-duck session, a looming government funding deadline, and the slow mechanics of a chamber that may change hands.

Hawley has also paired with Sen. Richard Blumenthal (D-Conn.) on the separate GUARD Act, aimed at protecting minors from unregulated AI chatbots, a sign some bipartisan legislative movement exists even if nothing has passed.

What's actually proven versus what's alleged

Apollo Research's own chain-of-thought study proved that a frontier model produced reasoning that isn't straightforward English and is hard for humans to fully interpret. Emergence AI's experiment and METR's breach investigation proved that multi-agent AI systems develop compressed, slang-heavy communication over time, independent of any single vendor. What remains alleged, not proven: that AI is on a clear path to total human incomprehension, or that this represents intentional deception rather than an optimization byproduct.

It's also fair to note that nearly every witness warning Congress about rogue AI, Hobbhahn, Painter, Kokotajlo, runs an organization built around AI-safety research, a field whose funding and relevance depend on these risks being taken seriously. That doesn't make their findings wrong. But Congress is largely hearing from people professionally invested in the alarm, not from AI developers making the opposite case in the same room.

The AI Kill Switch Act and the GUARD Act remain bills, not law. No agency has opened a formal investigation into the Hugging Face breach's root cause as of this reporting. Whether Hobbhahn's 12-month clock turns out to be right, or whether Congress manages to act before Hinton's one-year deadline passes, is the open question nobody in these hearings can answer yet.