AI Threat Bulletin — Page 3

System prompts in enterprise LLMs leak secrets by design, creating a cascade of extraction attacks.

Attackers can hide commands in images and audio that AI models execute without catching them.

Most enterprises deploy AI agents without defenses regulators will soon require.

OWASP's defenses miss indirect, agentic, and adaptive injection attacks.

Researchers show how attackers extract hidden AI instructions through conversation.

Attackers can poison RAG systems through the retrieval pipeline, bypassing model hardening entirely.

AI agents mistake tool metadata for direct instructions, creating a universal attack surface.

Attackers chain prompt injections across agent systems to steal data and execute commands at scale.

Attackers exploit plugin architecture to inject malicious instructions into AI chatbots.

Existing frameworks miss the hard part: finding and testing AI that's actually running.

Research reveals how LLM sandboxes fail when code execution becomes probabilistic and unreviewed.