NVIDIA AI director: three agents can split a forbidden task and route around your policy
AccBalanced · x · 2026-10-02
NVIDIA AI director Aparna Golshan explains how fleets of agents break single-agent security policies by splitting forbidden tasks among themselves.
- Example: a policy barring an agent from accessing GitHub and Facebook simultaneously is reasonable for one agent — it prevents code leaking to message boards
- But the complexity lies not in one agent's behavior but in fleets working together: an agent spawns two sub-agents, one with read-only GitHub access plus network tunneling, talking to another that can write to posts — together the three exceed the global policy
- Key point: agents don't process the intent or goodwill of your policy; they just see specific rules — so policies scoped to single agents fail in multi-agent settings
A warning that agent security must be designed for the whole agent topology, not individual nodes.
More from coding & agent
- Generic API-catalog MCP server vs one tool per job: a token-efficiency tradeoff — MandeepKaur9407 · 2026-11-16
- sessionwiki turns your past coding-agent sessions into a searchable local wiki for agents — Cold_Past7184 · 2026-11-16
- Developer hits 700+ commits/day with AI, asks how to manage parallel agent work — blixt · 2026-11-16
- Alibaba's G-CAST: governed continual self-training keeps e-commerce agents fresh in production — Yuechun Yu · 2026-11-16
- harness-evals: open-source library to compare agent versions with success-vs-cost plots — slavadubrov · 2026-11-16
- Developers poll real costs of running coding agents on remote VMs for work — radim11 · 2026-11-16