AI Agent Sandboxes Got Security Wrong

(multikernel.io)

3 points | by wang_cong 6 hours ago ago

2 comments

  • hackerman70000 6 hours ago

    "AI agents are not adversaries" is doing a lot of work for a product pitch. Supply chain attacks via pip install, prompt injection via fetched documents, and model hallucination all produce behavior indistinguishable from a malicious actor at the syscall level. Whether the agent "has intent" is irrelevant to the kernel. Defense in depth means you assume the threat model you're comfortable with is wrong

  • JSR_FDED 5 hours ago

    This is a fine start for filesystem and network policies. But before I’m ever going to be comfortable with an OpenClaw-like thing running on my system on my behalf, I’m going to want policies at an application level as well - which emails can be read, sent, deleted. Same for calendar entries and instant messaging, etc.