uri-eliabayev talks with tom-barnea, who leads AI-security detections at Tenable (exposure management; it acquired the AI-security startup Apex). AI risk differs from classic attacks. Instead of moving from a weak user to a big server, someone with one user’s access can simply ask a connected assistant for exactly the sensitive data they want. Because everyone can now build agents, many incidents involve HR, finance, or the business rather than the security team. Example: employees asking the company copilot about quarterly results right before publication, hinting at insider trading. Other risks: over-reliance (asking AI about chemical mixtures, medical questions, or student grades), self-promotional indirect prompt injection (hidden instructions in sites or résumés), and agents shared with too-broad permissions (one opened a user’s org chat history to the internet by misconfiguration). Vendor admin consoles offer far fewer controls for AI than for cloud, and none are aligned across vendors. Detection needs AI: semantic classifiers trained on labeled data, plus tightly scoped judge agents in production, built by security researchers paired with data scientists. Next: orchestrated multi-agent workflows, where each agent may be safe alone but the connections between them open new risk.
Show: hidden-layers
Concepts: agent-security · ai-guardrails · multi-agent-orchestration