AI Browser Agent Risk Checklist

It reads what you can't see, and acts with your session.

Security researchers at Brave demonstrated a working indirect prompt injection attack against Perplexity's Comet AI browser agent, hiding adversarial instructions in invisible page elements that caused the agent to execute sensitive cross-site actions — including fetching one-time passwords and accessing banking portals — using the logged-in user's own session and permissions. This checklist scores your own AI browser agent setup against the specific risk factors that made that attack possible: broad session access, unconfirmed purchases and credential autofill, and lack of visibility into what the agent actually did.

• →Built directly around the documented Brave/Comet indirect prompt injection proof of concept
• →Weights broad session access and unconfirmed purchases/credential autofill as the highest risks
• →Explains that AI agents read raw page content, including elements invisible to a human viewer
• →Covers action logging and a fast revoke/pause path for after-the-fact detection
• →Watermarked by Cikal Studio Labs · Works on any device, no install required