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.

I think if you let claude code go wild with auto approval something similar could happen, since it can search the web and has the potential for prompt injection in what it reads there. Even without auto approval on reading and modifying files, if you aren't running it in a sandbox it could write code that then modifies your browser files the next time you do something like run your unit tests that it made, if you aren't reviewing every change carefully.

I really don't get why you would use a coding agent in yolo mode. I use the llm code gen in chunks at least glancing over it each time I add something. Why the hell would you have an approach of AI take the wheel

It still keeps you in the loop, but doesn’t ask to run shell commands, etc.

That seems like a bad default. VSCode’s agent mode requires approval for shell commands every time by default, with a whitelisting capability (which is itself risky, because hiding shell commands in args to an executable is quite doable). Are people running agents under their own user identity without supervising the commands they run?