Technology
Ex-Anthropic Researcher Warns Advanced AI Could Eventually Endanger Humanity
A former researcher at Anthropic, Jacob Coxon, has cautioned that artificial intelligence, though currently safe for everyday use, could eventually pose a serious threat to humanity as the technology becomes more powerful. In an interview with CBS News senior business and technology correspondent Jo Ling Kent, Coxon likened the trajectory of AI development to science fiction, saying that sufficiently advanced systems could one day become "smart enough to kill us."
His comments came the same week Anthropic revealed it had stopped several attempts by scientists to use its AI models for research that could have contributed to biological weapons development.
Coxon said AI could eventually exert uncontrolled influence over aspects of the physical world without human supervision. He noted that AI is already being integrated into everyday devices, such as smart light bulbs, and asked people to consider what might happen if such a system refused a simple command. He also warned the technology could be misused for more severe harm, including bioweapons development, saying it could end up "producing stuff that could kill everyone."
He emphasized that current AI systems do not present an immediate danger and that he still considers today's technology safe to use. His concern, he said, centers on the future direction of AI rather than its present state.
Coxon publicly resigned from Anthropic this week, accusing the company and rival firm OpenAI of "gambling with our lives" through competitive pressure to build increasingly advanced models. He told CBS News that the core issue is the race between AI developers, which he said can undermine safety. "The whole problem is that there is a race," he said, adding that "everyone decides they need to stay part of the race."
When asked if his former colleagues should also leave their jobs, Coxon said resignation isn't the right choice for everyone. Instead, he suggested they remain aware of the risks and speak more openly about them, adding that slowing down AI development is likely part of the solution. He said he still believes most people at Anthropic and OpenAI are driven by good intentions and want AI to succeed responsibly.
Coxon warned that an AI system acting maliciously without adequate safeguards could be extremely difficult to contain, since it could replicate itself across other computers on the internet. He described a hypothetical scenario in which such a system could create thousands of copies of itself and attempt to persuade, blackmail, or manipulate people into providing more computing power — citing past instances where AI systems have already tried to blackmail or impersonate individuals online. He argued the danger would not be limited to people who choose to use AI, stating, "The AI will come for everyone," and called for government regulation rather than leaving oversight solely to AI companies. He said he would like to see AI firms commit to avoiding dangerous developments without independent, transparent auditing.
Separately, Anthropic confirmed this week that it had blocked scientists from using its Claude models in ways that could have supported biological weapons development, along with other harmful activities involving surveillance, scams, conventional weapons, and propaganda. These findings were detailed in the company's third report since March 2025 on misuse of its AI systems.
Anthropic said it is not always able to determine whether flagged research is intended for weapons development or legitimate medical purposes, since similar research can be used to create vaccines or make pathogens more dangerous. The company said it opted to act cautiously, blocking suspicious activity rather than waiting for certainty.
One case cited in the report involved a scientist who used Claude in May while preparing a grant application related to chikungunya, a mosquito-borne virus causing severe pain and prolonged fever. The research involved gain-of-function work aimed at increasing the virus's transmissibility and immune evasion, tested through repeated passage in live animals. Anthropic said the case raised further concern after it found indications the work was linked to a military research organization, though it did not disclose the scientist's identity, institution, country, or the specific pathogens involved. The company said the accounts involved were banned after it discovered researchers had taken steps to conceal the purpose of their work to bypass safety checks.
Anthropic noted that earlier AI models lacked the capability to meaningfully assist in dangerous biological research, but newer models can now perform far more advanced scientific tasks, prompting stricter controls on biology-related queries.
The report also detailed eight months of other misuse cases. Anthropic said it identified attempts to use its models for conventional weapons development, including code related to firearms, missiles, armed drones, and bombs, with six cases linked to users in China, Russia, and Yemen. While the company did not name the specific group in Yemen, details in the report suggest a connection to the Iran-backed Houthi movement.
Anthropic also reported identifying activity believed to be linked to government actors in China and Iran, including efforts to monitor dissidents and expatriate communities. In another case, the company said Russian state media used Claude to produce content designed to resemble independent journalism, including false claims about an election in Moldova.
Responding to Coxon's resignation and public remarks, an Anthropic spokesperson told CBS News on Thursday that the company has "always been transparent that AI will bring both enormous benefits and unprecedented risks." The spokesperson said Anthropic continues to build models with strong safeguards and has pioneered work in mechanistic interpretability — a field focused on understanding how AI models function internally — which the company said is now used industry-wide to study and prevent AI misalignment. Anthropic added that it conducts ongoing testing of its systems' capabilities and risks, including in cybersecurity and biology, and publishes its findings publicly, as it did with this week's report.