Back to Blog

They Didn't Build a Bioweapon. They Approved Their Own Change Ticket.

By: Casey Cannady : nomad, cybersecurity veteran & Chapter 7 survivor

November 16, 2026
11 min read
Casey Michael Cannady
CybersecurityAITechnologyPersonal Story

TL;DR

On September 18, Reuters reported that Anthropic, the AI company that has spent two years warning the world that its own models could help someone build a biological weapon, has quietly built a wet lab in the Bay Area. A real one. Benches, samples, physical experiments. The company confirmed it exists, said it is “not for drug discovery specifically,” and at first declined to say what it is for, how big it is, or what biosafety level it runs at. Five days later it published two of the three: fundamental biology research, at biosafety levels 1 and 2. It has still not said how big the lab is or who inspects it.

My first reaction was the same as yours: this is terrifying. My second reaction is the one I trust, because it is the one nearly 30 years in security trained into me. The lab is not the problem. Every biotech on earth has one. The problem is that a single company now builds the capability, grades the risk of the capability, sets the guardrails on the capability, and runs the bench where the capability touches the physical world. In my industry that has a name. It is a separation-of-duties failure, and we do not let a junior sysadmin get away with it.

Nobody hacked anything. Nobody broke a law. One company just wrote the change, approved the change, and audited the change, and asked you to trust the ticket.


I want to be honest about how this one started, because the honesty is the point.

I read the headline, and my stomach dropped. AI company. Secret-ish biology lab. Robots. Won't say what it is for. I fired off a message that said, more or less, this is slippery slope type shit.

Then I did the thing I tell everybody else to do. I stopped reacting to the label and went and read the permissions.

“Slippery slope” is the weak version of this argument. It asks you to be afraid of a future I cannot source. Regular readers know my rule from the gas-prices piece: I am not here to allege a villain in a smoke-filled room. I am describing architecture. And the architecture here is bad enough on its own that I do not need to invent a monster to put inside it.


What Actually Happened

Here is what is on the record, stripped of the spooky music.

  • The lab is real. Anthropic's head of life sciences, Eric Kauderer-Abrams, confirmed it to Reuters. His argument is simple: you cannot do biology entirely in a simulation, and the final test still happens at a real bench.
  • The purpose was fuzzy for five days. A spokesperson told Reuters the lab is “not for drug discovery specifically,” then declined to elaborate. On September 23, Anthropic said the lab's focus is “fundamental biology research using Claude.” By late October, Anthropic said it had begun drug discovery efforts of its own.
  • The spec sheet is half filled in. On September 23, Anthropic published the biosafety level, the single number that tells you what kind of work a room is physically built to handle: BSL-1 and BSL-2, with no pathogens that can infect humans. It said the research group formed in the spring of 2026. Still no size. No headcount. And no word on who inspects it. Its IPO filing now lists the lab as a business risk subject to biosafety rules, and still names no size and no inspector.
  • The robots are early, but real. One person familiar with the effort told Reuters the company wants to push how far Claude can direct robotic lab equipment with limited human involvement. The company's position is that human oversight stays essential, and it now says all of the lab work is performed by human scientists. Kauderer-Abrams describes AI-run lab execution as “very early innings.”
  • This is not a side project. Life sciences is now one of Anthropic's biggest areas by headcount and spending. In April it bought Coefficient Bio, a startup with fewer than ten employees, for roughly $400 million in stock. The same month, the CEO of Novartis joined its board. In June it shipped Claude Science, a research tool built for labs instead of chat. Genentech and Bristol Myers Squibb are customers, and Novo Nordisk signed on to test it two days before the lab story broke.

That is the whole known universe. Everything past that line is speculation, and I am not going to sell you speculation.


The Boring Truth About Wet Labs

Let me take some air out of the panic before I put the real problem on the table.

“Wet lab” sounds sinister if you have never worked near one. It just means a lab with liquids, samples, and physical experiments, as opposed to a room full of laptops. The overwhelming majority run at biosafety level 1 or 2: proteins, cell cultures, pipettes, a lot of plastic. The job postings Reuters dug up point at exactly that kind of work, protein and nucleic acid characterization, plus somebody to run procurement. Kauderer-Abrams himself says the setup looks like what most biotech companies do: some work in-house, some farmed out to partners.

So no, I do not think there is a supervillain lair in the Bay Area.

Here is what bothers me. One number would settle most of this, and for five days they would not publish it. Then they did: BSL-1 and BSL-2. A company that says “BSL-2, here is our scope, here is who inspects us” has ended the conversation. Anthropic has now said the first two. It has not said the third, and naming an inspector costs nothing. In nearly 30 years of security I have learned that when the cheapest possible reassurance is withheld, the withholding is the finding.


Four Hats, One Head

Here is the part that is actually my day job.

Every enterprise I have ever worked in runs on a rule so basic we teach it in week one. The person who writes the change does not approve the change. The person who approves it does not audit it. The admin who can push to 50,000 endpoints does not also get to sign off on what gets pushed. The person who cuts the check does not get to approve the vendor. We call it separation of duties, and it exists because of one hard-won lesson: good intentions are not a control.

Notice I did not say “because people are evil.” Most of the time they are not. We separate duties because a single party holding every key will eventually make a call that benefits itself, in good faith, under deadline, and nobody will be positioned to say no.

Now count the hats on one head.

  1. Anthropic builds the capability. By its own testing, its top model could give meaningful help to someone trying to build a biological weapon, help they could not easily get elsewhere. That is their finding, not mine.
  2. Anthropic grades the risk. The evaluations that decide how dangerous the model is are designed, scored, and published primarily by the company that sells the model.
  3. Anthropic sets the guardrails. It decides which biology questions get blocked and which get through, and it decides when to move that line. It also decides who gets a looser line: in late October it described a verification program that lets “trusted organizations” use its most powerful models with more permissive safeguards.
  4. Anthropic now runs the bench. The place where model output stops being text and starts being chemistry.

Four roles. One org chart. One revenue target. If I walked into a Fortune 500 and found one admin account that could write, approve, deploy, and audit, I would not ask whether the admin was a nice guy. I would write it up as a critical finding before lunch.


“We're Always Balancing These Things”

Reuters asked the obvious question: how do you square warning the world about AI bio risk with building out your own biology operation? The answer from Kauderer-Abrams was that “we're always balancing these things,” and that deployment is handled carefully.

I believe him. That is what makes it a problem.

“We are balancing it” is a self-issued badge. I spent a whole series on this blog showing you what happens to badges that are issued by the party they vouch for. The Signal phishers wore one. The fake agent skill wore one. Same shape here, minus the malice: the only entity telling you the balance is right is the entity doing the balancing.

You can watch it happen in miniature. Back in August, Anthropic retuned the biology safeguards on its top model. A lot of coverage called that “loosening,” and I want to be fair: the primary source says it was mostly a false-positive fix. The model had been refusing basic health questions, and the truly dual-use stuff (virology, toxicology, molecular design) still gets routed to an older, weaker model. Reasonable. Probably even correct.

But look at who decided. The company rewrote the rulebook for its own classifier, retrained it, tested it, and reported an 85% drop in blocked biology queries. Every number in that sentence was measured by the party being measured. I wrote almost this exact paragraph in the agent-skills post, about a security vendor grading its own breach demo. The standard does not change because I like the product better.


The Patch Exists. Check What It Covers.

Here is where I have to give credit, because this company is doing something its competitors mostly are not.

On September 12, Anthropic's CEO Dario Amodei published an essay called “We Must Pace the Frontier” and committed the company, unilaterally, to letting third-party evaluators inside with employee-like access. Desks, badges, laptops. On September 18, the same day the lab story landed, Anthropic named Accenture (its Faculty unit) as its first embedded evaluator, with each company committing at least a billion dollars over five years. On October 15 it added METR, a nonprofit, with six people given employee-like access, and said it is working with SecureBio as well. SecureBio later said publicly that Anthropic's description “did not reflect the agreement we have,” and Anthropic revised it.

That is a real step toward separation of duties. It is more than anyone else in this race has put on paper. Now let me do what I do with every patch: read what it actually covers.

  • An auditor with no veto is an observer. A banking-law scholar pointed out that the comparison to embedded bank examiners falls apart without the power to shut something down. The head of AI at security firm XBOW, whose team gets early access to unreleased models, put it more bluntly: “we don't have any veto power.” The final call stays with the company.
  • An auditor who is also your business partner has two jobs. Accenture sells Claude into enterprises. It is now co-investing a billion dollars alongside the company it is evaluating. I am not alleging anything. I am telling you what I would write on the whiteboard if a client proposed that arrangement for their own SOX controls.
  • Scope is everything. Everything I can find describes the evaluators testing safeguards, red-teaming models, and assessing alignment. I cannot find one public sentence saying the wet lab is in scope. Maybe it is. If so, saying it costs nothing.

This is the same lesson from the agent-skills post, wearing a lab coat. The half of the problem everyone can see is getting patched. The half that just got built is not in the patch notes.


It Is Not Illegal. That Is the Whole Point.

Before the guy in the comments says “so what law did they break”: none that I know of. That is the finding.

Biosafety oversight in this country leans hard on two hooks: federal funding and a specific list of regulated pathogens. A privately funded lab that takes no federal money and touches nothing on that list answers to a lot less than you would assume. There is no public registry where you can look up what a private company's bench is rated for. The rules were written for universities and government labs, in a world where nobody imagined a software company dropping $400 million on a biotech acquisition while shipping a model its own maker says can beat experts on some biology tasks.

So the disclosure is voluntary. The audit is voluntary. The human-in-the-loop promise is a company blog post, not a rule. Everything holding this together is a policy the company can rewrite on its own authority. And anything that depends on one party choosing to keep being good is not a control. It is a mood.


The Part Where I Have to Be Fair

I am not an AI doomer, and I am not a hater. I use Claude every single day. Parts of this post were drafted with it. When I asked Anthropic's own model to help me go after Anthropic, the first thing it did was disclose that conflict and tell me to check the sources myself. Which is, annoyingly, exactly the control I am asking the company for.

And I want the upside. I am a guy with a fused spine whose medical bills were a slow-motion bankruptcy before the legal one. The stated mission of the wider life sciences effort is rare and so-called undruggable diseases, the ones big pharma skips because the math does not pay. Amodei lost his father to a disease that became curable a few years too late. I do not doubt for one second that the people in that lab want to cure something.

I also notice who else is in the room. Reuters reported that Anthropic's own pharma customers may worry about the company learning from their competing drug programs, even with their data walled off. It sells the tool and it is building the bench. I sold software for a long time. I know what that arrangement is called, and it is not “partnership.”

Good people inside bad architecture is the most common breach report I have ever read. That is why we build the controls anyway.


I Refuse to Leave You in the Dark

Same split as always: the structural fixes that move the needle, and the personal moves that help you keep your head on straight. Do not confuse the two.

The levee (what we should be demanding):

  • Finish the spec sheet. Anthropic has published the biosafety level and the scope of work. Now name who inspects it, and make all three the standard for any AI company running physical biology. If it is boring, prove it is boring.
  • Put the lab in the auditors' scope, in writing. Embedded evaluators who can test the chatbot but not the bench are auditing the lobby.
  • Give evaluators a stop button, and fund them independently. An auditor who cannot say no and is paid by the audited is a consultant. Nonprofit and public evaluators with real authority, or it is theater.
  • Make human sign-off on AI-directed experiments a rule, not a press statement. Same for a registry of private labs where AI systems direct physical biology. The law was written before this existed. Rewrite it.

The sandbags (what you can do Monday):

  • Ask every AI vendor one question: who can say no? If the answer is “us,” you have your risk rating.
  • Apply the rule in your own shop. The agent that writes the code does not approve the deploy. The agent that drafts the payment does not release it. Separation of duties is not just for humans anymore.
  • Read the primary source. Some coverage flattened the August safeguards change into “they relaxed the guardrails.” The original tells a more complicated story, and it is usually one click away.
  • Refuse both scripts. “AI will kill us all” and “trust us, it is fine” are the same request: stop checking. Do not.

Sandbags are sandbags. Your personal vigilance does not audit a lab in California. Only rules with teeth do that, and only organized people get rules written.


Why I'm Telling You This

Because the fear is the easy part, and the fear is useless.

If you walk away from that headline terrified, you will do what terrified people do: nothing, or something stupid. If you walk away reassured because a nice man said they are balancing it, you will do what reassured people do, which is also nothing.

I am asking for the third thing. The boring thing. The thing that has kept every system I ever defended standing: nobody, no matter how smart or how sincere, gets to write, approve, and audit their own change. Not a sysadmin. Not a bank. Not a company that, by its own published testing, holds a capability it considers dangerous, and just built a room to plug it into.

Despair says the machines are coming and there is nothing you can do. Fury says this is a governance gap with a known fix, written down in every security framework on earth, and the only thing missing is someone with the authority to enforce it.

They did not build a bioweapon. They approved their own change ticket. Go ask who signs the next one, and get loud with me.


Sources & Further Reading

Sourcing note: the facts about the lab come from Reuters' reporting, the outlets that carried it, and Anthropic's own September 23 post. The biosafety level (BSL-1 and BSL-2) is the company's own statement; I could not find an outside party that has confirmed it. The separation-of-duties argument is mine, built on nearly 30 years of doing this for a living. I could not find any public statement on whether the wet lab falls inside the embedded evaluators' scope; if Anthropic says it does, I will update this post and say so loudly. Disclosure: I use Anthropic's products daily, including in drafting this piece. Verify anything you plan to repeat.


Connect with Casey

Have a story, a question, or want Casey to write about a specific topic? DM me and tell me which story you want next.

Casey writes about economic policy, nomadic life, cybersecurity, and navigating the world as a late-diagnosed AuDHD adult. New posts drop on my professional website.