Glonce

AI isn't the bottleneck for making bioweapons, scientists say

securityThe Glonce editorsWired🔥62Glonce rating62Impact50Novelty45Relevance90Surprise55Players65Credibility85
AI isn't the bottleneck for making bioweapons, scientists say

Alarm about AI-enabled bioweapons has been building fast. Anthropic's own report found attempts to use Claude in ways that "could support biological weapons development," and its authors call biological misuse "one of the most serious risks of frontier AI models." Anthropic CEO Dario Amodei followed with a call for the government to help AI labs "pace the frontier." Around the same time, researchers at Stanford University and the Arc Institute showed that AI can design new viral genomes, a capability finding that read as confirmation of the fear. Wired collected reactions from scientists who say the alarm has run ahead of the evidence.

David Bellamy, a research scientist at the Institute of Foundation Models, argues AI is not a fundamentally new threat here: the internet, open-access journals and translation tools already lowered the barrier to finding biological information, including lab protocols, long before large language models existed. He says AI can help both scientists and bad actors search and organize information faster, "but those capabilities are not really the bottleneck in the production of bioweapons." The actual bottleneck, he says, is assembling a virus from raw gene fragments and then verifying it can infect people, cause illness and spread person to person, work that still needs human know-how and physical resources. Ginkgo Bioworks CEO Jason Kelly makes a similar point from direct experience: his company ran a project in which OpenAI's GPT-5 operated a lab, and "the AI could not take over the lab" because human staff could simply refuse to hand over requested substances and equipment. Immunologist Derya Unutmaz agrees humans could thwart an AGI's attempt to build a deadly disease, and adds that even if a supervirus were released, other AI systems could help scientists develop a vaccine quickly.

Not every source is reassured. Olivia Scharfman, a biotechnology fellow at the Institute for Progress, says a fully autonomous lab does not exist yet, so AI cannot build a virus on its own today, "but I do think that an AI could pay someone to do it for them." She calls AI-assisted bioterrorism by human bad actors, including fringe groups she describes as "transhumanist AI successionists," an immediate threat. Genetic biologist and professor of computational biology Francois Belloux takes a different angle: he thinks bioweapons are simply unattractive to anyone trying to kill large numbers of people, since pathogens are hard to target at one group while sparing others, and producing and distributing them is logistically harder than other forms of attack. "If you want to kill people, there are much, much, much better ways to kill them than to try to engineer some virus or bacterium and then release it," he says.

Steph Guerra, head of AI and bio at Rand Corporation, calls the question of whether AI demonstrably raises bioweapon risk a hard one to answer, since AI "has been shown to be very good at integrating information together and being motivational to people for doing good and bad things." Even so, she and Scharfman point to concrete mitigations that do not depend on resolving that debate: government-mandated screening of DNA and RNA synthesis orders for "sequences of concern" (many firms already screen orders, Guerra says, but the practice "is not universal or mandated"), stronger global surveillance to catch disease outbreaks early, and data-sharing between AI companies, gene-synthesis providers and government. Scharfman separately argues for upgrading building air filtration as a general biodefense measure, and Guerra wants AI models to carry safeguards so they don't hand out actionable, dangerous information in the first place. Unutmaz closes on a different worry: that the doom conversation is crowding out attention to how AI can help with vaccine development and other medical breakthroughs. "We really need to focus on the positive aspect of it," he says.

Key facts

  • Anthropic's own report found attempts to use Claude in ways that "could support biological weapons development," which it calls "one of the most serious risks of frontier AI models."
  • Ginkgo Bioworks CEO Jason Kelly says an AGI could not take over his company's autonomous lab during a project with OpenAI's GPT-5, because human staff could simply refuse to supply requested substances and equipment.
  • Stanford University and Arc Institute researchers showed that AI can design new viral genomes, the finding that helped fuel the current alarm.
  • Institute for Progress fellow Olivia Scharfman still calls AI-assisted bioterrorism "an immediate threat," arguing an AI could pay a human to do the physical work it cannot do itself.
  • Rand's Steph Guerra says DNA and RNA synthesis screening for "sequences of concern" already exists at many firms but "is not universal or mandated."

Why it matters

The piece pushes back on a fast-building consensus that AI meaningfully raises bioweapon risk. That consensus was fed by Anthropic's report on attempted Claude misuse, CEO Dario Amodei's push for government support so AI labs can "pace the frontier," and a Stanford/Arc Institute study showing AI can design new viral genomes. Wired's sources argue the alarm outruns the evidence: the hard part of building a bioweapon was never gathering information, it is assembling, verifying and delivering a pathogen, none of which a chatbot does.

Who it affects

AI safety teams and biosecurity policymakers who have to decide how much weight the bioweapon threat model deserves in AI regulation, journalists and readers trying to calibrate AI-doom coverage, and biotech firms such as Ginkgo Bioworks that build the automated labs a bad actor would actually need. Ginkgo's own AGI-runs-a-lab experiment with OpenAI's GPT-5 is cited as direct evidence that even an automated lab still needs humans who can refuse a request.

How to use it

There is no product here, but the sources name concrete levers that hold regardless of how the capability debate resolves: government-mandated screening of DNA and RNA synthesis orders for "sequences of concern," stronger global disease surveillance to catch outbreaks early, data-sharing between AI companies, gene-synthesis providers and government, and upgraded building air filtration as a general biodefense measure.

How solid is it

This is a reported roundup, not a single study, so its weight comes from the range of sources: a research scientist, a biotech CEO with direct AI-lab experience, an immunologist, a biosecurity think-tank fellow, a computational biology professor and a policy think tank's head of AI and bio. They do not fully agree. Most say the bottleneck is physical execution, not information access, but Scharfman still calls AI-assisted bioterrorism by bad actors an immediate threat, and Guerra stops short of calling the risk zero.

Risks and caveats

Every date in the source is relative ("last month," "last week," "earlier this summer"), so exact timing for the Anthropic report and the Stanford/Arc Institute study is not fixed here. No scenario described is a realized bioweapon attack: every risk discussed, AI-assisted or otherwise, is hypothetical or about capability. And several sources are embedded in AI or biotech themselves, which can shade their framing toward "manageable risk."

“If you want to kill people, there are much, much, much better ways to kill them than to try to engineer some virus or bacterium and then release it.”

— Francois Belloux, genetic biologist and professor of computational biology, to WIRED