Why Prompt-Level Guardrails Fail for AI Agents
summary
Claim: prompt-level guardrails fail once autonomous agents can run tools, read files, adapt and self-modify, so sandboxing or microVMs on a laptop is insufficient. Demonstrated vectors include NPX pulls that can exfiltrate AWS keys and agents that move binaries or create symlinks to bypass command lists. Proposed defenses are kernel-level boundaries enforced with eBPF/Falco, tamper-evident kernel audit logs and a continuous observe layer feeding an adaptive runtime security loop that hardens posture and triggers countermeasures.
AI-written summary · about
chapters (6)
- 0:00Why prompt-level guardrails fail
- 3:17Autonomous agents change the attacker model
- 7:27Why a laptop sandbox is not enough
- 12:10Kernel boundaries the agent cannot edit
- 20:16Adaptive runtime security from kernel data
- 28:28Practical defense in depth
related talks
17mGuardrails First: Designing AI Agents for Security Operations
25mPlease Sandbox your AI Agents
11mBuilding Secure Backend Services with AI Agents
30mSandboxing Agentic AI With LSM-BPF
24mAI Governance for developers: A practical framework for deploying autonomous coding agents safely…
34mSecure Your Coding Agent Like It’s Malware
39mBuilding Safe AI Agents
19mKeynote: Towards Building Safe and Secure Agentic AI