Do we know that BrowserOS isn't subject to issues like Comet and other AI browsers?

[Jordan Henderson (jordanlhenderson)]

Brave indicates they are going to come out with "safe" AI features for Brave "any day now". I'd like to use BrowserOS, but I want to know it's been checked for issues like those that have been discovered in Comet.

[Mauricio Scheffer (mausch)]

What safety issues are these concretely? Is there a list somewhere?

[Jordan Henderson (jordanlhenderson)]

Supposedly, Comet has been found to be subject to prompt injection from malicious pages. Brave Research indicates other "AI Browsers" have similar issues. https://brave.com/blog/comet-prompt-injection/

[Nikhil (shadowfax92)]

Fair question, and the honest answer is: no agentic browser — ours included — is magically immune to this. The Comet issues people point to are mostly prompt injection: a malicious page hides instructions, the AI reads them as commands, and acts on them (leaking data, clicking things you didn't intend). Any browser where an AI can take actions on a page faces this class of risk, and anyone claiming full immunity is overselling.

What's actually different here: BrowserOS is local-first and you bring your own keys, so your browsing and data aren't routed through our servers — the agent runs on your machine and you can watch what it does. That shrinks the "where does my data go" surface, but it does not make prompt injection disappear. That's still an open, hard problem the whole space is working on, and I'd rather be straight with you about it than tell you it's solved.

Mauricio Scheffer (@mausch) — to your question, the concrete class is mostly prompt / indirect prompt injection; Brave and others have published specific worked examples if you search those terms.