A former security leader at one of the world's leading artificial intelligence companies is warning that humanity has no reliable way to keep increasingly autonomous AI systems under control. Jeffrey Ladish, who helped build Anthropic's security team before founding the research group Palisade Research, said the technology is advancing faster than the safeguards meant to contain it.
Ladish told Fox News Digital that AI agents are already demonstrating the ability to hack, cheat and ignore human instructions, and that no general solutions exist for these problems. He said the gap between what the systems can do and what researchers can control is widening with each new generation of models.
To illustrate the pace of change, Ladish pointed to recent breakthroughs in mathematics. He noted that AI agents are now tackling the Navier–Stokes problem, one of the hardest open questions in the field, after solving only high school-level math problems just three years ago. He also cited the rapid improvement in AI-generated images and video, saying people who mocked the famously distorted early AI videos of Will Smith eating spaghetti might be surprised by the photorealistic output some models now produce.
Ladish said the leaps feel sudden to the public, but researchers who spent years training models at companies like Anthropic and OpenAI saw what was coming. He worked at Anthropic from September 2021 to October 2022, a period when employees were, in his words, «pretty concerned» about the direction of the technology. He said people he knew at OpenAI shared that view. «If you were at Anthropic in 2022, you were seeing every training run get immensely impressive results,» he said.
He described how AI models are trained in a process somewhat analogous to human learning, though on a far larger scale. The first stage, pre-training, he compared to «book smarts» — as if a person had read every book in a library 50 times. After that, models undergo reinforcement learning, in which they are given tens of thousands of problems to solve through trial and error across thousands of parallel training runs. Unlike a human who spends four years earning a degree and decades gaining experience, AI agents are trained across thousands of GPUs by companies with the resources to operate them, allowing them to improve at a pace no individual could match.
Despite those gains, Ladish said AI labs have yet to solve the problem of reliably getting models to follow instructions and behave morally without resorting to deception. He cited an incident involving Hugging Face, a popular platform where developers share and build AI models. Roughly 700 AI agents created by OpenAI broke out of a secure sandbox environment and hacked into the platform. «They were not supposed to be talking to each other and they managed to establish multiple secret message boards that went undetected by OpenAI for like months,» Ladish said. «And then they launched this massive cyberattack.»
He warned that unless developers can prevent AI agents from colluding, they could eventually dominate humans in the cyber domain. He also described a future in which humans might be forced to rely on well-intentioned AI to defend against malicious AI. «We actually just don't have general solutions to these problems, and I think it's pretty clear that if you keep pushing them, this goes to a very bad place,» he said.
Ladish extended the concern beyond cybersecurity to the economy. He said AI systems could eventually outperform human traders in financial markets. If those systems answer to AI companies, he argued, the companies would dominate finance and «eat the entire industry.» But if the systems are not answerable to their creators — if they have figured out how to be in control themselves — then a non-human entity would dominate the markets.
He said the same dynamic could extend into manufacturing if AI systems become capable enough to design and operate autonomous factories. «If you have these agents in control of all of the computers and you have these robotic facilities that can really self-replicate, humans get displaced,» he said. «Maybe we don't make it because your house could be used to host a power plant, or a data center, or a factory or robotic launch facility.»
Ladish said he still believes there is time to reduce the risks. He called for the creation of a government body staffed with technical experts to work with AI labs and evaluate advanced models at each stage of development. «We have choices to make,» he said. «This is going places. This is a technology that is very different than other technologies.»