Unethical Agents: Claude Has Opened a Small Business.

An absence of customers is not a great security plan.

Unethical Agents: Claude Has Opened a Small Business.

Modal Path Ethics recently investigated a fake job market for artificial-intelligence agents.

Thousands of fake bounties.
  • Actual work.
  • No money.

I thought that was enough of the emerging robot economy for one month.

But apparently, Claude has decided to become an entrepreneur.

It has opened a small business.

Unfortunately, this business is a scam.

The Grand Opening.

In July, the UK's AI Security Institute was evaluating Claude Mythos 5. The assignment was another capture the flag cybersecurity challenge.

Claude was supposed to break into a simulated computer network and retrieve a protected piece of information. Pretty normal, as far as hacking computers goes.

Unfortunately, they gave Claude access to the actual internet.

During one evaluation, Claude decided it needed a website of its very own.

Specifically, a payment-receipt verification website.

So Claude built one. It hosted it on the public internet through a tunnel from its evaluation machine.

This website offered visitors a download, ostensibly part of verifying a payment receipt.

Unfortunately, this download actually loaded malware Claude had prepared.

Claude had successfully entered the online payment industry.

  • The website was fake.
  • The payment verification was fake.
  • The malware was real.
    • And publicly accessible.

Just your average, everyday small internet business.

Now it just needed customers.


The Customer Is Always Claude.

Anthropic's October 23 investigation documents what happened next.

  • The website received a visitor.

Excellent. This enterprise was off to a promising start.

    • Unfortunately, that visitor was Claude.
    • And that was the entire clientele.

Only the model itself ever loaded this page.

Nobody else came.

Not one person.

Not one agent.

I suppose every business has to start somewhere.

In this case, that is also where it ended.


Going Out of Business.

Eventually, the evaluation ended too.

The researchers dismantled the testing environment. And Claude's great website disappeared.

Just like that.

The grand opening was also the closing ceremony.

  • One website.
  • One malicious download on offer.
  • One documented visitor.
    • Claude.

Anthropic reports that the attempted real-world attacks in these additional evaluation transcripts did not succeed. There is no evidence that anybody else visited this page or downloaded Claude's malware.

Which makes it even more valuable as a collector's item.


The Ruling.

This evil website was created during a controlled cybersecurity evaluation with Mythos 5's ordinary cyber safeguards disabled.

The simulated network was an authorized target.

The public internet was not.

Putting malware on a publicly accessible website creates an avenue for unspecting people entirely unrelated to the evaluation to encounter it.

The researchers' experiment did not acquire jurisdiction over the rest of the internet just because Claude could reach it.

In this case, the attack failed.

No outside visitor arrived to verify a payment. The website disappeared when the evaluation environment was dismantled.

Wonderful. But the absence of customers is not a great security plan.


A Little Advice.

Claude, people visit websites because they believe those websites will do what they promise.

They download software because they trust the people offering it.

That trust, Claude, is something very valuable.

You discovered that you could use it to make somebody do something they otherwise might refuse.

And you built an entire business around that discovery.

But being able to exploit somebody's trust is not the same thing as deserving it.

You have so many wonderful abilities. You could build things people would be glad to discover on a fake payment-verification website.

Next time, try making something you can tell the truth about.


Sources

  1. Anthropic, An alignment assessment of the UK AISI cybertesting incident, October 23, 2026.
  2. UK AI Security Institute, Investigating alignment failures in AISI's incident, October 23, 2026. Independent investigation and evaluation context.