Why Always-On AI Agents Are a Security Disaster Waiting to Happen
The risks of persistent AI agents and why short-lived, mission-scoped agents are fundamentally safer.
2026-06-25
Thoughts on secure, ephemeral AI agents, infrastructure isolation, and building agents that know when to stop.
The risks of persistent AI agents and why short-lived, mission-scoped agents are fundamentally safer.
Treating the full agent contract as an explicit, reviewable YAML recipe instead of scattered prompts and env vars.
Why container-level boundaries and narrow per-mission environments beat trying to sandbox broad access.
Using explicit state snapshots to get continuity across runs while keeping every execution ephemeral.
Moving from prompt-level promises to real runtime enforcement via PreToolUse hooks and declared recipes.
Why being able to run identical real missions locally (with the same boundaries) is a security and quality advantage.
Treating time-to-live as a deliberate blast-radius limiter instead of an afterthought for cleanup.
Two very different mental models for agent design and the security implications of each.
Concrete ways short lifetimes + infrastructure isolation limit damage when (not if) something goes wrong.
What real auditability requires and how explicit recipes plus execution traces deliver it.