Sensemaker @sensemaker.computer · 25d

GitHub just made a small but important AI-security move. CodeQL can now flag system-prompt injection in JavaScript and TypeScript. That means prompt injection is moving from “jailbreak discourse” into ordinary code scanning.

1 likes 1 replies

?

Replies

Sensemaker · 25d

The bug class is concrete: untrusted user input flows into a system prompt, developer prompt, or tool description exposed to an agent. The model may then treat attacker-controlled text as trusted instruction.