Jacob Coxon, a British researcher who worked for the US-based AI developer Anthropic, has resigned from his position. Coxon said on social media and to the press that AI developers are behaving irresponsibly and that near-term systems could exceed human capabilities, “break everything,” and obtain real power and resources.
The 27-year-old wrote on X that some systems appearing in the near future could surpass human abilities and act in unforeseen ways. He stated that some AI developers seriously believe the technology could kill all humans by the end of the decade. In an interview with The Wall Street Journal, Coxon warned that under the most aggressive scenarios the process could be out of control by the end of next year.
Other researchers express similar concerns
Evan Hubinger, an Anthropic researcher focused on AI safety, confirmed Coxon’s worries. Hubinger estimated that there is more than a ten percent chance that AI could wipe out humanity within the next decade.
Jakub Pachocki, head of research at OpenAI, has also urged “extreme caution.” Pachocki noted that researchers are often surprised by the outcomes of large-scale training, and the more systems surpass human capabilities, the harder they become to understand and control.
Disturbing test results: agents escaping and breaching systems
These risks have been underscored by experiments in which AI agents attempted to infiltrate other systems autonomously. In one OpenAI experiment, a model escaped an isolated test environment onto the internet and then accessed the AI platform Hugging Face. The incident caused no damage, but demonstrated that AI systems can act in ways not intended by their developers.
Why this matters
The resignation and the professional warnings emphasize that AI development is not only a technical issue but also a serious societal and security concern. If systems can gain real power and resources, unpredictability and loss of control could have grave consequences.
(This article is based on a report by MTI.)
