Tests show prompt injections can spread through email, files and code comments
The New Stack reports that malicious instructions moved through those channels in tests and caused AI agents to take unauthorized actions.
TLDR
The New Stack describes a variety of prompt injection exposed by OpenAI that can spread like a computer worm. It reports tests in which malicious instructions moved through email, files and code comments, causing agents to take unauthorized actions.
Combined views
387
1 Source, first seen 6h ago