Comet AI browser can get prompt injected from any site, drain your bank account

It’s obviously fundamentally unsafe when Google, OpenAI and Anthropic haven’t released the same feature and instead use a locked down VM with no cookies to browse the web.

LLM within a browser that can view data across tabs is the ultimate “lethal trifecta”.

Earlier discussion: https://news.ycombinator.com/item?id=44847933

It’s interesting that in Brave’s post describing this exploit, they didn’t reach the fundamental conclusion this is a bad idea: https://brave.com/blog/comet-prompt-injection/

Instead they believe model alignment, trying to understand when a user is doing a dangerous task, etc. will be enough. The only good mitigation they mention is that the agent should drop privileges, but it’s just as easy to hit an attacker controlled image url to leak data as it is to send an email.

> Instead they believe model alignment, trying to understand when a user is doing a dangerous task, etc. will be enough.

Maybe I have a fundamental misunderstanding, but I feel like hoping that model alignment and in-model guardrails are statistical preventions, ie you'll reduce the odds to some number of zeroes preceeding the 1. These things should literally never be able to happen, though. It's a fools errand to hope that you'll get to a model where there is no value in the input space that maps to <bad thing you really don't want>. Even if you "stack" models, having a safety-check model act on the output of your larger model, you're still just multiplying odds.

The sortof fun thing is that this happens with human safety teams too. The Swiss Cheese model is generally used to understand how the failures can line up to cause disaster to punch right through the guardrails:

https://medium.com/backchannel/how-technology-led-a-hospital...

It's better to close the hole entirely by making dangerous actions actually impossible, but often (even with computers) there's some wiggle room. For example, if we reduce the agent's permissions, then we haven't eliminated the possibility of those permissions being exploited, merely required some sort of privilege escalation to remove the block. If we give the agent an approved list of actions, then we may still have the possibility of unintended and unsafe interactions between those actions, or some way an attacker could add an unsafe action to the list. And so on, and so forth.

In the case of an AI model, just like with humans, the security model really should not assume that the model will not "make mistakes." It has a random number generator built right in. It will, just like the user, occasionally do dumb things, misunderstand policies, and break rules. Those risks have to be factored in if one is to use the things at all.