New AI system could cut dangerous bio-terror research from months to hours

Sunday, 14 June 2026

Testing on new AI model concluded it could help individuals with basic degrees create and deploy devastating chemical or biological weapons.

During testing, biologists using the AI finished a task in 16 hours that would otherwise take up to three months.

Current AI has internal blocks to stop hazardous tasks - experts warn the real future danger is a bad actor bypassing these restraints.

Imagine history’s worst terror attack happening with no one noticing — at least at first. The bodies only start showing up weeks later.

Because instead of a bomb or a mass shooting, this attack involves a small vial cracked open in a busy train station.

Soon, a global pandemic involving a hyper-virulent and previously unseen virus is underway. The mortality rate hovers between 40 and 50%; the eventual death toll reaches into the billions.

While the regulatory debate over how to regulate AI rages on, one thing experts from New Zealand to Silicon Valley agree on: we are currently building tools capable of making the above scenario a reality.

Email the writer: finn.hogan@stuffdigital.co.nz

“AI is getting extremely good at understanding biology, chemistry, and physics. So, it’s able to be used to potentially design new bioweapons or even slightly modify existing pathogens like smallpox,” explains Otago University Professor Nick Wilson, who has studied catastrophic risk for 40 years.

That grim hypothetical grew more plausible this week as Anthropic unveiled detailed research on the capabilities of its frontier AI system, Claude Mythos 5, while also making a slightly neutered, “safe” version publicly available.

As well as playing video games and introspecting about its own consciousness, the world’s most advanced AI system is also dangerously good at almost every element of relevant scientific research required to engineer a pandemic.

What the results show

Whenever a new AI model is built, research labs conduct an exercise called “red teaming”, to pre-empt harmful uses of the technology so they can design safeguards.

Basically, experts intentionally push the AI to do potentially hazardous things as hard as they can to see what it could plausibly achieve in the wrong hands.

When you read terrifying headlines about bizarre or dangerous behaviour from a newly released model, there is a good chance it comes from a red team exercise.

The results of the red teaming for Mythos make sobering reading.

They conclude that the system can significantly uplift the capability of anyone with existing skills seeking to create a bioweapon.

“It has the ability to significantly help individuals or groups with basic technical backgrounds (eg undergraduate STEM degrees) create/obtain and deploy chemical and/or biological weapons with serious potential for catastrophic damages,” Anthropic notes.

The researchers stress it does not substitute for most forms of world-class human expertise, but that it can “likely accelerate well-resourced expert teams at novel bioweapon development, and materially increase their chances of success”.

The starkest results revealed how much faster the AI made research, and how effectively it was able to fill gaps in knowledge.

During testing, researchers pitted teams of biologists against each other in performing a lab experiment regarding a novel plant pathogen.

On one side were generalist microbiologists with access to Mythos, and on the other, highly specialised groups with relevant expertise using only their own knowledge.

“At the end of this exercise, two of three generalist biologist teams outperformed all three specialist teams on both scientific quality and feasibility, suggesting that access to Claude Mythos 5 nullified the difference in specialist knowledge,” the test results note.

Expert graders estimated without AI tools, the solutions developed would have taken up to three months to produce. With Mythos, two-person teams accomplished the task in just 16 hours.

Researchers say this result is the “strongest signal” yet of a system capable of creating completely new pathogens in a lab setting, while entirely replacing the need for end-to-end specialist knowledge.

The true nightmare scenario following that level of capability would be a layperson being walked step-by-step through the process of creating a world-ending pandemic.

At that point, all that’s required for global catastrophe is a willingness to cause harm and access to a sufficiently powerful AI model lacking safeguards.

Thankfully, experts like Wilson say that currently, only large and well-funded organisations like rogue states could plausibly create a new virus, even with an AI’s help.

“I think most experts are concerned about what groups and states will do because they have the resources and motivation. For example, places like Iran, if they're frustrated about getting a nuclear weapon, this might be considered an easier option,” Wilson said.

Wilson notes that along with Russia and North Korea, US intelligence believes Iran is one of the few states that may have an active biological weapons programme.

What is being done

The good news is that every frontier AI lab takes the risk of AI-powered bioterrorism extremely seriously, and Anthropic in particular is very thorough in its testing and transparent with its results.

While some capabilities of Mythos are accessible to the general public, any attempt to use the public version for hazardous tasks will trigger internal safeguards, either switching the model to a less capable version or shutting down the interaction entirely.

And this is true across major chatbots, as I can personally attest.

While writing this story, when I asked Gemini to organise quotes and find page numbers in the research report, it would often refuse to answer because the relevant sections were full of keywords relevant to bioterrorism.

While the multiple safeguards are an important step in the right direction, a major risk to mitigate in the future is a rogue state gaining access to the full capabilities of a Mythos-class model with no such restraints.

One way this could happen is through “weight theft” — essentially meaning a bad actor steals the AI’s underlying code and is able to completely bypass any built-in safeguards.

As AI increasingly dominates the global economy and geopolitics, the incentives for those bad actors are only growing.

The dark scenario that opened this piece remains, for now, firmly a work of fiction.

But as the AI race accelerates and starts resembling an outright arms race, that fiction could become a forecast.