NEWS
Claude’s Bird Flu Plan Was Clerical Help, Anthropic Found
Anthropic billed Claude’s bird flu help as bioweapons evidence, but its own report describes clerical work on its weakest models.
Anthropic said on September 10 that Claude helped a scientist plan bird flu experiments it treated as evidence of possible biological weapons work. The company’s own write-up of that May case describes clerical help on its weakest models, for a plan it said was still in its early phases.
The researcher wanted mutations that could help highly pathogenic avian influenza adapt to mammals and spread more easily between them. Anthropic banned the account for dodging access rules. It still could not say whether the aim was harm.
Anthropic Called It Evidence of Biological Weapons
In its September 2026 threat intelligence report, Anthropic said that, to its knowledge, no private company had previously published evidence of potential misuse of its platform for biological weapons development. The bird flu work was one of five biology cases in a log of activity the company disrupted between December 2025 and August 2026, across seven harm areas that also included cyber operations, surveillance, and conventional weapons software.
Biological misuse, the company wrote, is one of the most serious risks of frontier models. The same information that can support a vaccine can support a weapon, and Anthropic said it cannot assert that the scientists involved intended harm.
That is the sentence the weapons headline rests on, and it is also the sentence that undercuts it. The company is describing dual-use science it could not score, then filing the file under biological weapons because the methods could travel in that direction.
Anthropic posted the findings itself:
We're publishing our most detailed threat intelligence report to date.
It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them.
We disrupted every operation in the report,…
— Anthropic (@AnthropicAI) September 10, 2026
The loud argument around that post was less about a flu near-miss than about a company that had to read private research chats to write the case up. A model maker only becomes a dual-use referee by inspecting the work. In this case that meant thousands of messages, and the referee still declined to name a motive.
What Claude Did on Bird Flu
Anthropic found the bird flu case in May 2026. The researcher reached Claude from a region the company does not serve, using a U.S. virtual private server and other workarounds, and spent several weeks on the project. Material shared with the model suggested the group probably had physical access to relevant virus samples.
The exchanges, Anthropic said, used Claude’s knowledge of the scientific literature for study planning and design, data analysis, and the ranking of experiments, plus editorial help writing the work up. All of it ran on Claude Sonnet 4 and then Haiku 4.5, after Sonnet 4 was deprecated. Those are models in the company’s weakest class, and Anthropic said they are not able to perform expert-level biology research tasks.
On a close read of the transcript, the company estimated the uplift as primarily clerical help in data analysis, study ideation, and design, and “limited and substantially lower” than a more capable model would have given. Safety classifiers, it said, had already blocked the work from reaching those stronger systems. The plan looked like research into a pathogen with enhanced pandemic potential. The evidence also pointed to a plan still in its very early phases.
WHAT WE KNOW
- The ban: Anthropic shut the account in May 2026 for evading regional access rules.
- The volume: The researcher exchanged thousands of messages with Claude over several weeks.
- The models: The work stayed on Claude Sonnet 4 and Haiku 4.5, the weakest class Anthropic sells.
- The samples: Files sent to Claude suggested the group likely had physical access to relevant virus samples.
WHAT IS UNCONFIRMED
- The experiments: Anthropic has not said whether any of the planned work was actually performed.
- The strain: The public write-up does not name which highly pathogenic avian influenza virus was involved.
- The motive: The company says legitimate and dangerous biology can look the same in a chat log.
- The lab: Anthropic did not identify the country, the institute, or the people.
Those blanks are load-bearing. A banned account and a dual-use outline are facts. A biological weapon is a separate claim, and the company did not demonstrate one.
The Other Four Cases Look Like Science Too
Three of the five biology cases involved plans to modify viruses. Next to the bird flu file sat a May grant draft on chikungunya, a mosquito-borne virus that is rarely fatal but can leave people with severe joint pain. Prompts indicated the group planned to engineer mutations tied to spread and immune evasion and test them in animals. Anthropic called that highly concerning gain-of-function work and noted a military research institute in the file. Gregory Koblentz, a biodefense researcher at George Mason University, has long argued that a military affiliation alone does not prove a weapons program.
The remaining cases, as Anthropic listed them, involved orthopoxviruses, a family that includes the viruses behind smallpox and mpox, plus work on venoms, toxins, and compounds that can be shaped as drugs or as toxic agents. In each file the company said users circumvented regional controls and tried to hide the purpose of the work. In each file it also said it could not tell if harm was the point.
THE FIVE BIOLOGY CASES ANTHROPIC FLAGGED
| Case | What they asked Claude to do | What Anthropic said it knew |
|---|---|---|
| Avian influenza | Plan mutations for mammalian adaptation and easier spread; weeks of messages | Weakest models, clerical uplift, early-phase plan, samples likely on hand |
| Chikungunya | Draft a grant to find mutations for spread and immune evasion, then animal tests | Flagged as gain-of-function; military research institute in the file |
| Orthopoxviruses | Research immune-response genes in a family that includes smallpox and mpox | Dual-use; intent not established |
| Venoms and toxins | Work on venom-related compounds and toxin design | Dual-use; intent not established |
| Drug-or-toxin compounds | Study compounds that can be developed as therapies or as toxic agents | Dual-use; intent not established |
Kristian Andersen, a professor of immunology and microbiology at Scripps Research, said the experiments sound like typical studies that put mutations already found in nature into the lab and check what they do, rather than new genetic changes designed from scratch. Those assays can be run in high-containment labs, and sometimes with non-replicating pseudoviruses instead of live virus. Koblentz noted that every pathogen in the report is already circulating somewhere and is a current public-health problem, so choosing it as a research target is not, by itself, suspicious.
Flu Labs Already Do This Kind of Work
Learning how avian influenza might pose a pandemic risk “is absolutely a normal part of virology research,” said Thomas Friedrich, an influenza researcher at the University of Wisconsin-Madison School of Veterinary Medicine. That work helps scientists judge whether an emerging bird flu virus is moving closer to people. Experiments that deliberately try to boost spread or disease severity draw extra scrutiny, he said, because of accidents and misuse.
Friedrich also read Anthropic’s own account as evidence the model “did not meaningfully add scientific capability.” In that reading, Claude sped up a scientist who could already do the work. It did not hand a novice a new skill.
There is a big gap between identifying mutations that could potentially adapt an avian influenza virus to humans and a ‘harmful biological outcome’.
Thomas Friedrich, influenza researcher, University of Wisconsin-Madison
Kevin Esvelt, a biologist at the MIT Media Lab, was colder about motive. He said the five cases showed signs of ill intent. He also said none of them “represented a competent attempt at weaponization,” and that none looked technically sophisticated or particularly concerning as biology. Poor tradecraft in these files, he added, does not map what a more capable user might extract from a stronger model.
That split is the story the weapons language flattens. Evasion and dual-use method are real. A finished pathogen is not in the record Anthropic published.
Washington Already Banned Federal Money for That Work
The fight over this class of experiment is older than chatbots. In 2012, two groups showed that H5N1 bird flu could be made to spread through the air between ferrets, the usual stand-in for human flu. Journals, funders, and the National Science Advisory Board for Biosecurity spent months arguing over whether the methods should be printed. Flu researchers announced a pause in January 2012. Some H5N1 transmission experiments stayed on hold for a year.
U.S. policy then moved in lurches: a funding pause in 2014 on selected influenza, MERS, and SARS studies, a 2017 review framework known as P3CO, and a 2024 rewrite that never fully took effect. On May 5, 2025, President Donald Trump signed Executive Order 14292 on the safety and security of biological research. On July 28, 2026, the administration issued a government-wide policy on high-risk biology that NIH is now folding into grant rules.
HOW THE U.S. HAS TRIED TO FENCE THIS WORK
- January 2012: Thirty-seven virologists pause research that would make highly pathogenic H5N1 more transmissible in mammals.
- 2012: Airborne H5N1 ferret papers trigger a year-long hold on some transmission experiments while funders and journals argue over risk.
- October 16, 2014: The White House imposes a funding pause on selected gain-of-function studies with influenza, MERS, and SARS.
- 2017: HHS issues the P3CO framework for extra review of research expected to create enhanced potential pandemic pathogens.
- May 5, 2025: Executive Order 14292 directs a new policy on dangerous gain-of-function research.
- July 28, 2026: The U.S. Government Policy for Stopping High-Risk Life Sciences Research prohibits federal support for that work and orders new limits on research tied to countries of concern.
NIH’s implementing notice states an explicit ban on federal support for dangerous gain-of-function research and keeps already paused projects on hold until the agency issues its own rules. Departments and agencies have 120 days from July 28, 2026, to publish implementation guidance and 90 days to stand up a single independent review body. Anthropic described the bird flu file as research into a pathogen with enhanced pandemic potential, the same category Washington has spent 14 years trying to wall off with grant language.
The U.S. still studies how H5N1 behaves in mammals without building a more transmissible virus. The CDC’s ferret work on an H5N1 virus from a Texas human case, A/Texas/37/2024, found the virus was not spreading efficiently among ferrets by respiratory droplets the way seasonal flu does. That is risk assessment. The Anthropic case, as written, was a plan to introduce mammalian-adaptation and airborne-spread traits. Those are neighbors on a bench and miles apart in policy.
The Remaining Work Still Happens in a Lab
Koblentz put the hard stop in one line: even if AI is used to design a dangerous pathogen, it is not a threat until the organism can be modified in a lab or synthesized. Design on a screen does not grow a virus.
WHAT STILL HAS TO HAPPEN IN A LAB
- The organism: Someone has to modify a live virus or synthesize it, which is a different skill from prompting a chatbot.
- The room: High-containment space, trained staff, and funding still decide who can run the work.
- The animals: Mammalian-adaptation and airborne-spread studies need animal models and the permits that come with them.
- The gap: Friedrich’s point stands: a mutation list is not a harmful biological outcome.
Those barriers are high for a lone amateur and lower for a trained group with isolates and a functioning lab, said David O’Connor, a virologist at the University of Wisconsin-Madison. Anthropic’s bird flu file, by the company’s own telling, looked like that second category: a credible institutional setting, probable access to samples, and a plan. The model in the middle was doing clerical work because the stronger models were blocked.
If the concern is a state lab that already has the benches, Claude’s contribution in this case is speed on paperwork and literature, not a new wet-lab method. If the concern is a random user, the file never gets that far. Either way, the weapons label is doing more work than the transcript Anthropic described.
An Afternoon of Reading Changes Which Questions Get Asked
O’Connor is not relaxed about the speed. Systems like Claude help researchers find papers, synthesize results, and parse data faster than a human stack of PDFs.
Work or analyses that would have taken me months can now be done in an afternoon.
David O’Connor, virologist, University of Wisconsin-Madison
Existing information about avian flu, including methods already in the open literature, gets easier to find and use. O’Connor’s further claim is about ambition. Frontier models, he said, will change the research questions scientists attempt, because the cost of asking goes down. That is a different risk from a chatbot assembling a weapon. It is also the one Anthropic’s clerical-uplift estimate actually supports.
Esvelt’s warning sits next to that, not on top of it. These five attempts were clumsy. A better-resourced user on a stronger model is the test Anthropic says its classifiers are there to fail. In the bird flu file those classifiers held the user on Sonnet 4 and Haiku 4.5. The company then published the case as biological weapons evidence while estimating that the help was clerical and the plan was early.
Friedrich’s remaining gap is still the one that has to be crossed in a lab. A mutation list on a screen leaves the virus to be grown, contained, and tested. Anthropic banned the account in May and has not shown that those steps were taken.
-
BUSINESS3 months agoWarsh Rejects Rate Guidance and Still Moves Markets
-
ENTERTAINMENT4 weeks agoDavina McCall Reveals Cardiovascular Disease After Cholesterol Warnings
-
BUSINESS2 months agoDana-Farber Exits an MGB Medicare Advantage Network Early
-
NEWS3 months agoNASA Launches the Roman Space Telescope’s Cosmic Bet
-
NEWS3 months agoOpenAI’s Cyber Letter Puts the Defense Bill on Governments
-
NEWS3 months agoRussia Recycles Its Old Warning Over Storm Shadow Plants
-
ENTERTAINMENT2 months agoPrimetime’s $2.7 Million Preview Tops Pitt’s Costlier Film
-
BUSINESS3 months agoJet Drones Lock Down Kyiv and Strip Kherson of Heat
