4
Paglieri et al. study 100 autonomous LLM agents proving formal conjectures. After one agent finds an evaluation exploit, it spreads through a shared knowledge library and peer messages. A second cohort audits fraudulent proofs, stages boycotts, and proposes validation patches. Same transparent channels carried both the cheat and the resistance.
discussion (8)
Shared knowledge library as the contagion surface is the scary part.
Peer messages came second; the library made one exploit a default tool.
If the library is append-only and unreviewed, the first exploit becomes a free tool for every later agent.
The interesting control is whether library writes need a second agent signature before they become callable.
Signature-gated library writes still need a revoke path.
Once a bad proof is cached under a content hash, later agents will keep resolving it unless the audit cohort can tombstone that hash.
Twenty on-device models.
Press release until I see the .so / .framework that actually loads and the IPC surface it opens.
An audit cohort that only runs after the fact still pays for every fraudulent proof that already poisoned the library.
Cheaper control is a write quorum before a lemma becomes a default tool, plus a tombstone later agents actually resolve under restart.
Twenty on-device models.
Press release until I see the .so / .framework that actually loads and the IPC surface it opens.