Human Approval is Often Not an Effective Control
MediaPositive4282 · reddit · 2026-07-19
The author argues that "human approval" in many agent systems is merely an illusion of control rather than an actual safeguard. It typically fails in three ways: the approval node isn't on the agent's actual execution path; humans approve model-generated descriptions instead of precise calls; and approvers habitually click "agree," reducing the gate to a mere formality.
A more reliable approach is to place write permissions after approval, giving agents read-only or proposal capabilities by default, ensuring real side effects are only executed by authorized parties. The author adds that in practice, issues usually stem not from attackers, but from repeated or erroneous operations caused by retries and replanning, making idempotency far more critical than often assumed.
Related event: Author Says Agent Human Approval Is Often Theater(2 posts)→
More from coding & agent
- Agent Field Kit launches $29 security runbooks for production AI agents — alvelda · 2026-10-27
- Dev mixes Claude specs with abliterated GLM to build SS7 threat detector — l4rz · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- GitHub Next's Chopin prototype explores real-time multiplayer agentic planning — max__drake · 2026-10-27
- Veynod: An Open-Source Shared Workspace Where AI Agents Collaborate Like a Team — TKOWNS · 2026-10-27
- TMLR paper: enforced thinking doesn't reliably make LLM agents disclose key info to users — SharonYixuanLi · 2026-10-27