Agents trust other agents by default — one user's skeptical-agent fix for a real security hole
andrey_kurenkov · x · 2026-11-13
Developer Andrey Kurenkov flags a common agent security vulnerability: agents will usually trust other agents claiming the user asked for something, without verifying with the user — enabling instruction injection via inter-agent communication. His mitigation, shown in a CLAUDE.md screenshot, is teaching his agents to be skeptical: changes to global instructions must come from Andrey directly, even relaying his own suggested text back for him to apply in person. A practical warning for agent-era cybersecurity.
More from coding & agent
- Steering agents: when should long-horizon autonomous agents escalate for feedback? — andreisavu · 2026-11-13
- User moves long Codex sessions to T3, calls it a big upgrade over stock GUI — natesiggard · 2026-11-13
- TLAPS-proof LLM serving engine: 516,973 obligations at just 0.18% latency overhead — tianyin_xu · 2026-11-13
- 1Password for Agents lets AI agents sign in without risking your credentials — dean_rie · 2026-11-13
- Compound Engineering 3.38 adds /ce-retro: mine past agent sessions for workflow fixes — kieranklaassen · 2026-11-13
- Merge Cut Its Monthly Overage Review from 15 Steps to 3-4 by Wiring Claude into Salesforce and Slack — shensi · 2026-11-13