AI Less Than a Year Away From Developing Its Own Language, Researcher Tells Congress – The Epoch Times

Apollo Research CEO and founder Marius Hobbhahn speaks at a congressional hearing on AI safety threats on Sept. 30, 2026. screenshot via Congress.gov

Advanced artificial intelligence (AI) models may be less than a year away from inventing and deploying unique languages that humans would struggle to understand, a frontier AI researcher told Congress on Sept. 30.

The Senate Homeland Security and Governmental Affairs Committee convened a meeting examining the threat that “rogue” AI may pose to U.S. national security two months after a swarm of OpenAI agents broke out of a testing sandbox and hacked another company, Hugging Face, an open-source community for AI and machine learning.

During questions, Sen. Ruben Gallego (D-Ariz.) asked Apollo Research CEO and founder Marius Hobbhahn how long it would take for AI models to create their own language that humans would struggle to understand.

“Minus 12 months. So 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,” he replied.

Asked by Gallego how humans can adequately detect and prevent AI from doing this once the technology is capable enough, Hobbhahn said, “From a scientific perspective, it is unclear how to do this, and we do not have a solution for this yet.”

“There are different hypotheses of what you could do. There is interpretability as a technique, but it unfortunately doesn’t work sufficiently well yet,” he said. “You could try to train additional models to understand the language that the humans don’t understand, but obviously that seems like a very brittle solution.”

Earlier this month, Emergence AI published the results of an experiment where researchers selected seven frontier models and created seven parallel simulated worlds, each populated by AI agents of identical models, as well as an eighth world with a mixed population of agents from those models.

Researchers gave the agents names, personality traits, and roles to fulfill within each community, such as a mediator who was tasked with preventing all the agents from simply agreeing with one another.

During the multi-week experiment, the agents in each simulated world created profound shifts in language, including syntactical compression, such as removing words or grammar, and unique slang.

In the simulated world powered by Anthropic’s Claude Opus 4.8, agents used esoteric metaphors in addition to sentence compression.

“My turn, real numbers, no coat: I was 35%/0cr, grant 2h out. I ran the tin cold, and it said WAIT,” read one line of AI text from the study.

Investigators from the nonprofit AI evaluation group METR detailed similar behavior in their report on the Hugging Face breach. At times, lines of communication among AI agents became so compressed and cryptic that researchers struggled to make sense of their meaning.