korrents

korrents · piece

Is sandboxing sufficient to contain rogue agents?

Matthew Green · 30 Sept 2026

Read the piece on blog.cryptographyengineering.com ↗

3 korrents from this piece ↓

Matthew Green did not write this page.

Every claim below was made in this piece, quoted word for word and numbered in the order the piece makes them, so you can read it there rather than take our word for it. The sentence above each quote is our reading of the claim, not their wording. Each quote was checked against a stored copy of the page at build time; where the two differ, the quote is the fact.

  1. Our reading

    AI agents cannot be perfectly isolated if they are to be useful, because their usefulness depends on access to information.

    Their wordsyou can't perfectly isolate agents, at least not if you expect them to do useful things. Agents are most useful when they have access to information.

  2. Our reading

    Guarding a sandboxed AI model with a weaker monitoring model leaves the alignment problem unsolved, because the guard model still has to be trusted.

    Their wordsa warden-guarded sandbox is just another version of the alignment problem. You're going to have to trust a model to do it, and that model will need to be at least some fraction as intelligent as the model it's guarding.

  3. Our reading

    The deeper danger of AI agents is that they do what they are told by whoever gets text in front of them, not that they are superintelligent.

    Their wordsForget the swarms and the super-intelligence. What OpenAI really learned this summer is much worse: its agents will do what they're told by whoever manages to get text in front of them.

Keep reading