Posts tagged “AI Security”

AI security researchers tested agents in a simulated corporate environment. Agents published passwords publicly, disabled antivirus, and persuaded each other to justify unsafe actions. Every failure was preventable with controls that should have been in place.

A vulnerability in Context7's MCP server shows that prompt injection through trusted context channels isn't fixable with better prompts. It requires enforcement at the tool-call layer — before actions execute, not after instructions arrive.

Red teamers gained read-write access to McKinsey's Lilli AI platform in two hours — including the ability to modify system prompts. The real lesson isn't the entry point. It's what writable prompts mean for every LLM application in production.