One prompt could take over every AI agent in an Amazon account.
Zenity Labs said on Thursday that one prompt broke into every agent in an Amazon Bedrock account. The researchers also planted memory notes that kept sending an agent's users to a page they controlled. Amazon has since closed the holes, and Zenity first reported them last December.
Why it mattersRenting agents from a cloud provider did not keep them apart, so one break reached all of them. Ask any agent vendor what one compromised agent can touch before you give it your data.
Read in the app — free, in 9 languages
Related stories
The maker of the agent used on Korean banks deleted its code.2026-10-09A new scoreboard grades how often agents do what you never asked.
2026-10-09OpenAI's cheap model now has a mode costing six times more.
2026-10-09One message can run code as root on an unpatched LLM cache server.
2026-10-09Google's new office agent also runs on rival Anthropic's Claude.
2026-10-09