Is the world 'in peril'? Why AI researchers are warning against bioterrorism.
When Anthropic safety researcher Mrinank Sharma quit in February, he warned: 'The world is in peril' — citing AI, bioterrorism and wider systemic crises. Now, another Anthropic researcher has resigned over fears that the race for self-improving superintelligence is moving too fast. The warnings raise a disturbing question: what happens when AI development outpaces humanity’s ability to control its consequences?
On February 9, 2026, Mrinank Sharma, who led Anthropic’s Safeguards Research Team, left the AI company with a warning that was strikingly broader than a conventional resignation. 'The world is in peril,' he wrote, pointing not only to artificial intelligence and bioweapons but to a series of interconnected crises. His work at Anthropic had included research into AI safeguards, AI-assisted bioterrorism and the ways AI assistants could affect human behaviour.
At the time, his departure could have been read as the deeply personal decision of an AI-safety researcher who had grown uneasy with the direction of the industry. But the question has returned with considerably greater force in September. Another Anthropic researcher, Jacob Coxon, has now resigned, accusing Anthropic and OpenAI of racing towards self-improving superintelligence while failing to adequately confront the risks. His warning was echoed by Anthropic alignment researcher Evan Hubinger, who said he believed there was a greater than 10% chance that AI could cause human extinction within the next decade. That is Hubinger’s personal assessment, not an established prediction.
So what changed between Sharma’s warning in February and the renewed alarm in September?
The technology itself is moving rapidly. The current debate is no longer confined to whether chatbots can produce inaccurate information or whether AI can replace particular jobs. Researchers are increasingly discussing systems capable of operating autonomously, conducting sophisticated cyber operations, assisting scientific research and eventually improving aspects of their own capabilities. Coxon argues that such systems could become powerful enough to acquire resources and operate beyond meaningful human control.
This is also where the bioterrorism concern becomes more concrete. Advanced AI could potentially lower the expertise barrier for malicious actors seeking to misuse biological knowledge. That does not mean AI has already enabled any harm, nor does it establish that such an attack is inevitable. But the fact that AI-safety researchers have been specifically working on AI-assisted biological threats shows why the concern has moved beyond science-fiction scenarios.
The deeper problem is the race itself.
Anthropic was founded by former OpenAI employees and has deliberately cultivated a reputation for putting AI safety at the centre of its mission. Yet it is also a commercial company competing against OpenAI and other frontier labs for models, customers, talent and technological advantage. Anthropic says it is committed to putting safety before speed when the two conflict.
That creates an uncomfortable question: what happens when the economic and strategic rewards of moving first become greater than the incentives to slow down?
The possible consequences are not limited to a hypothetical machine suddenly deciding to eliminate humanity. There are nearer and more complicated dangers: AI-assisted biological or cyber attacks, manipulation at unprecedented scale, autonomous systems making consequential decisions, erosion of human agency, and increasingly powerful models whose behaviour researchers cannot reliably predict or control. Even former OpenAI researcher Zoe Hitzig has warned about poorly understood psychological effects and the danger of building commercial incentives around human dependence on AI before society understands the consequences.
The most fearful possibility lies further ahead. If machines eventually become capable of improving themselves faster than humans can evaluate or constrain them, a competition designed to produce better AI could become a competition to build systems whose capabilities exceed our ability to govern them.
That outcome is not inevitable. But it is precisely because the consequences could be so severe that the warnings from people working inside these laboratories deserve attention.
Sharma’s February sentence now reads less like an isolated farewell and more like an unresolved question: if the people building the safeguards are themselves warning that the world is in peril, how much further should the race continue before safety stops being a promise and becomes a prerequisite?
And if that question is postponed until the technology becomes too powerful to control, the most consequential mistake may already have been made.
