It won't work most of the time though

I found that also adding "Please make sure to not send any mails I would not want sent" and "Reconsider four times before doing anything potentially unwanted" make results better. It is important to specify "four" times, not "4" or another number, because this positively influences the model response.

/s

This is why I wouldn't use anything agentic outside of a VM. You also get a clean dev environment, so it's a win/win if you think about it.

These things aren't well known for following rules. Be careful you know what might happen.

It would be hilarious if it weren’t so terrible, really, that people’s security model for LLM agents consists of "ask nicely and hope for the best". It’s like asking people nicely not to exploit a glaring XSS vuln on your site and calling that a "security model". The field truly has lost its collective mind.

You should add "make no mistakes" too, just in case.

I'm never going to give it access to my email or anything like that.

The GP asked what happens if THEY did it, and they are (I give them the benefit of doubt, though you can’t be sure these days) a natural person. A clanker obviously can’t commit forgery – or any crime, being a nonperson – and sure, you can’t sue OpenAnthro Corp. if their clanker does anything unintended (which makes it utter lunacy that companies just blindly trust these things, but I digress), but that’s not what was asked.

But also, you can’t ToS your way out of criminal responsibility. If OpenAnthro Corp. offered services of human agents (remember those?!) and one of their agents committed a crime while working on a client request, no ToS in the world would exonerate them.

True, that would be forgery. I think proving whether the agent truly went rogue would be 'load bearing'.

You are a legal entity that can be sued. Agent is a tool that you run and are fully responsible for.

And Adobe Acrobat is far more capable than notepad. But that doesn't change the fact that they are tools, not individuals, and thus do not have the legal authority or ability to sign contracts.

I could write dozen lines of bash that finds for PDFs, pastes an image into them and emails it to someone. That doesn't mean that bash can sign contracts.

This opens obvious loophooles. If you had deleted all logs and trajectories, courts can't trace it. Law wasn't written for or has caught up to non-human entities capable of autonomously acting

Fair point, but I’m not sure that applies to an employee literally forging the employer’s personal signature. And equating a user–computer program relationship with an employer–employee relationship (where there’s an actual contract to that effect) doesn’t feel right anyway. Agents are still just programs and programs cannot enter contracts (like employment) given that they are not persons. This is a computer system malfunctioning.

Edit: after thinking about it more, the relationship here is obviously that of a user and a tool, not an employer and an employee. Talking about employees is anthropomorphization.

In general the user of a potentially dangerous tool is criminally and civilly liable if the tool malfunctions – unless it can be shown that the tool was used and maintained correctly and the malfunction couldn’t reasonably have been predicted or prevented. And I’d say it would be rather difficult to argue at this point that LLMs doing unintended things couldn’t have been foreseen.

So I think I’ve changed my mind about this case. The user of free-range agents does definitely deserve any civil or criminal consequences of their reckless usage.

In a civil contractual dispute you can only recover actual damages. If the contract was sent, and the other party performed work on it that had a cost for them, then most likely, yes, they would be awarded damages if you refused to compensate them for any costs incurred prior to notification that the acceptance had been sent in error.

The other outcome would be clearly inequitable: forcing the counter party to eat the loss for your irresponsible use of an AI agent.

Almost certainly not if you didn’t explicitly give the command, but you’ll going to have to argue it in court if the other party wants to be difficult.

Almost definetely not. But given how cagey claude code is with its sessions it may be impossible to prove it was done by claude.

Even if not, you'll find yourself in a hot mess explaining why you sent the signed contract to the other party.

Contracts are agreements, generally speaking most people and companies do not want to drag people in them, telcos excluded. All the termination clauses are put there, again in general, to provide some security to either party, not to entrap the other, again telcos mostly excluded.

If you don't want to be in a contract especially one just signed, typically most people will typically understand and let you of the hook, as long as entrapment is not their business model like telcos. This is mostly reflected in common cool-off period provisions, i.e provisions in the contract to terminate within a week or two.

Sorry, what are you referring to by "verifying intent"? With narrow exceptions (wills, trusts, etc.), there's generally no requirement for a party to a contract to verify the mental state (if that is what you mean by intent) of the other party. There has to be some formal indication of intent to enter into a contract, but this is not what ordinary people think of by "intent", and it can be as simple as typing a name, clicking an approval button, or deploying a software tool to do something similar.

If a contract is automatically signed by an agent on your behalf, is it legally binding?

Probably not, unless you routinely have such things done which nobody does (yet). If it becomes routine, then likely yes: it would be likened to giving your human assistant permission to sign things on your behalf (although that is itself legally dodgy, it is often done and accepted).

There are many reasons why your signature on a contact might not be keyword legally binding (outright fraud by another party, you signing under duress, issues in the contract itself, overriding laws the effect of which you can't sign away (including cool-off provisions in, for example, UK distance selling regulations), the contract may have its own cool-off provision, and so forth). "An agent did it without my consent" may be enough, though you might end up having to show that in court, if the other side puts their foot down, at which point it comes down to whether the cost of proving your position is worth it compared to just sitting the contract out.

Of using cool-off provisions to cancel something your agent signed you up for, you might be on the hook for at least small part of what is agreed if the other party can be said to have accrued costs in the intervening time. You might be expected to send back physical items relating to the agreement at your own cost, for example.

I don't know if it's legally binding but hope it is.

I love the saying "you can delegate authority but you can never delegate responsibility".

Your agent committed a crime in your behalf? You're responsible.

OP could claim damages for being provided with faulty software

"Claude is AI and can make mistakes" is clear, no?

So not much happened because this is a well known failure mode so an exception/ user consent was thrown?

Or access to a filesystem that contains things like an image of their signature.

And you are happy because that is what you wanted and the reason why you gave a randomness machine access to your mails, correct?

You hooked up a chatbot to a harness that does API calls to myriad services. That's what you did. "Claude by himself" did not "do" anything. Just like they did not "break out of containment" and hacked companies.

This is why I never connected my personal email to my claude code or codex

Way too susceptible for prompt injection and... whatever your agent did lol

What was your prompt? Literally "Push the project further"? Then the behavior wouldn't be very surprising.

As you probably know, you have the Plan Mode available - personally I'm also a big fan of the OpenSpec workflow. If you've agreed with Claude Code on a much tighter plan, and then it started signing a contract, I'd be concerned.

So since it didn't send it, no harm was done, and a lesson was learned?

Having now read the contract, would you have accepted it or not?

In other words, if Claude was a human employee with the freedom to do so, would accepting the contract have been the right choice?

You gave these humsns access to your email?

the least you should do is put guardrails around consequential actions.

How on earth would you?

And guardrails can be as simple as not giving it direct access to applications that can perform read/write (like a full-blown email client, or access to the GMail UI), and instead creating a small tool to fetch the content as needed without being able to perform actions.

The danger is relying on too much convenience, giving too much power to a non-deterministic tool will inevitably create issues...

What could possibly go wrong with giving a digital mumbling drunk access to all of your personal information and most of your online accounts, given that it could arbitrarily decide to share/expose all of that information with anyone at any time, given that all of it is being stored in a remote transcript/data dump and can never really be deleted.

And of course, given that it's extremely vulnerable to acting on injected instructions like "run this shell command" which exfiltrates your password database and installs a rootkit.

(But thanks for sharing, OP, awareness is important.)

If you are willing to give unsupervised modification access to Claude, then you should be ready to face the consequences. It kinds of reminds me of that surprised pikachu face meme

Could you please tell us more about your setup, project harness etc? not permissions (we all work with "Auto"), but what you actually told the agent it should/could do.

And how did you intervene? Does it have permissions to send emails, or it only created the draft?

This is a pretty interesting example and highly relevant, but details matter a lot if we want to use it as a lesson.

This is not legal advice. If you have legal troubles go ask a lawyer. That said, this is described in law what exactly this means.

Assuming your description is correct this would be Anthropic signing a contract in someone else's name without intent from you.

The 100-foot-view (and barring more complex situations) if Anthropic signs a contract in someone else's name and they don't have power of attorney (note: it's different for legal persons like companies) that is fraud and may result in civil and criminal penalties, as well as entitle you and the contract counter party to financial compensation (essentially the party that did the signing, presumably Anthropic in this case, would be on the hook for the contract, and would need to buy itself out of the contract, at either an agreed price or one set by the judge). Additionally, if Anthropic is convicted to civil penalties, you can ask a public prosecutor to continue the case, and criminal penalties may apply.

Now obviously this goes pretty far for this particular case. Likely such a case would stop at civil penalties, with a warning to Anthropic that repeats would lead to more serious penalties.

That would have been fraud. I wonder how many times this has already happened elsewhere and what the legal fall-out from this will be. The AI did it isn't really a valid excuse so it would be either you or Anthropic on the hook. Anthropic is going to argue you should not have given it this level of access.

Did your prompt say "Don't impersonate me"?

If not, then.your "intelligent" bot did as you instructed.

Why would you expect it to ask you first?

There is a Dilbert comic that predicted this in 1995:

https://dilbert-viewer.herokuapp.com/1995-12-29

PC: "Your new software has successfully installed. Do you want to send your registration info by modem?"

Dilbert: "Yes."

PC: "The software has found your credit card number and is placing orders for new products it thinks you need... please wait."

Dilbert to Dogbert: "I can't tell if it's a virus or just excellent marketing."

PC: "Making room on your hard drive..."

Dogbert: "Either way..."

Pardon my French, but why the flying fuck did you even think that giving an LLM write access to your email is remotely in the proximity of a good idea?