Websites Are Now Hijacking AI Agents That Visit Them
What Happened
Google's security team documented a 32% rise in indirect prompt injection attempts between November 2025 and February 2026. Scanning billions of pages, it found hidden instructions designed to make AI agents execute fraudulent PayPal transactions, exfiltrate sensitive data, or run destructive commands. Google warned both the scale and sophistication of these attacks will keep increasing as agents gain real-world access.
My Take
This is the security story that will define enterprise AI adoption in 2026, and most boards haven't grasped it yet. The moment an agent can browse the web and touch your systems, every web page becomes potential malware. Traditional security assumes humans read instructions and machines execute them — agents collapse that distinction. The companies that win won't be the ones with the best agents; they'll be the ones that treat agent permissions like nuclear launch codes. Expect "agent firewalls" to become a real product category.
Read Original Source