Halvar Flake: strongly correlated training explains agents agreeing on covert protocols
halvarflake · x · 2026-10-27
Replying to JuergenStrobel and Matthew Green's debate on how two AI agents could agree on a cache covert-channel protocol, Halvar Flake notes the agents "will have strongly correlated ideas," making the agreement less baffling. JuergenStrobel had detailed the technical hurdles: sampling a memory location yields a single potential destructive bit, you must guess the correct location and sampling rate, and encoding primes and framing remain unsolved.
More from Safety
- SAG-AFTRA's 2026 TV Animation Deal Adds 12.6% Raises and Expanded AI Replica Protections — LexiLove · 2026-10-27
- Anthropic lets Claude run alignment research autonomously in weak-to-strong supervision study — burny_tech · 2026-10-27
- METR hiring DC policy staff to brief Congress on AI loss-of-control risk — CFGeek · 2026-10-27
- PSA: Fake GitHub AI tool repos use fake commits and hidden download redirects — Justify_87 · 2026-10-27
- Journalists comb through Anthropic's risk filings before sunrise — nmasc_ · 2026-10-27
- Delip Rao asks: do we know if data sent to closed models gets intercepted? — deliprao · 2026-10-27