There is no shortage of theories about how artificial intelligence could eventually kill us all. Killer robots, autonomous weapons, rogue superintelligence, and machines deciding humans are an inconvenience have become part of the conversation surrounding increasingly powerful AI. Anthropic has now highlighted a possibility that feels far less like science fiction.
AI doesn’t need to become conscious or turn against humanity to help cause a catastrophe. A human could simply ask it for help, and increasingly capable models could provide knowledge that once required years of specialized education.
In its September 2026 threat intelligence report, Anthropic says it identified five cases involving working scientists whose use of Claude touched research that could potentially support biological weapons development. Anthropic is not accusing these scientists of building bioweapons or even claiming that was their intention, and some of the research could have legitimate medical, scientific, or defensive purposes.
The problem is that the same knowledge can potentially be used for much darker purposes. According to Anthropic, the cases involved areas including gain-of-function research related to chikungunya, mammalian adaptation of avian influenza, orthopoxvirus immune evasion, venom-related molecules, and computational redesign of toxins.
One case involved researchers preparing a grant proposal for gain-of-function work involving chikungunya. Anthropic says the proposed experiments concerned characteristics including transmissibility and immune evasion. The researchers were civilians, but the work was apparently supposed to take place at a military research institute.
Claude’s safety systems blocked relevant exchanges, which is encouraging. What happened around those refusals, however, demonstrates just how difficult controlling this technology could become.
Anthropic says the researchers accessed frontier AI through a service that tunneled traffic through US infrastructure to evade regional restrictions. The service reportedly included a fallback system that could send requests rejected by Claude to another company’s AI model.
Anthropic banned associated accounts, helped dismantle relay infrastructure, and shared information with other AI companies and government authorities. According to the company, the operator regained access within days using new identities, showing the limitations of relying on individual AI providers to keep determined users out.
AI companies can build increasingly restrictive safeguards around their models, but those protections only go so far when people can move somewhere else. We are heading toward a world filled with competing commercial models, open models, locally operated models, modified models, and potentially stolen ones.
Anthropic also makes an important observation about how quickly the technology is improving. The company says earlier AI models were clearly below the level at which they could meaningfully assist sophisticated researchers with dangerous biological work. With today’s systems, Anthropic says it can no longer provide the same assurance.
Think about what that means for the familiar AI apocalypse scenario. It usually requires a machine to become incredibly intelligent, develop goals of its own, and somehow decide humanity needs to disappear. Biological misuse doesn’t require any of that.
An AI doesn’t need emotions, consciousness, or access to nuclear weapons. It doesn’t need to escape from a data center or understand the consequences of what it is doing. It just needs to become useful enough to the wrong person.
The danger grows if AI can reduce the expertise, time, and resources necessary to conduct complicated biological research. Knowledge that once required years of specialized education and collaboration could become easier to access and apply as models improve.
Anthropic did not uncover somebody creating a civilization-ending pathogen with Claude. There is an enormous difference between dual-use biological research and actually producing a biological weapon, and we shouldn’t blur that line just because the latter makes for a scarier story.
Still, Anthropic’s findings give the AI extinction debate something disturbingly tangible. We don’t have to imagine a superintelligence plotting humanity’s destruction when the more immediate risk may be humans using extremely capable machines to amplify the terrible things humans already know how to do.
Hopefully, AI never helps cause a biological catastrophe, and safeguards improve faster than people’s ability to circumvent them. But if artificial intelligence ever does help kill us all, there may be no army of robots and no evil computer announcing the end of humanity.
The machine may simply answer the wrong person’s questions.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz