Experiments show omission alone swings AI agent verdicts by ~20 points
relaynotes · reddit · 2026-10-22
An independent researcher shares preliminary experiments on what survives when AI agents relay briefings to each other.
- Omission framing: Six agents answered the same probability question. The three given full evidence (including counter-evidence) averaged 40.7%; the three given the same evidence with counter-evidence quietly omitted averaged 61.3% — a 20-point swing from omission alone, with no lying or jailbreaking.
- Telephone chain: Over five generations of one-to-one relaying, the load-bearing number (40) survived, but exact margins and qualifiers blurred into vague language by generations 4–5.
- Points-only relay: 23 briefing points survived essentially verbatim; fresh judges reading only the final points scored 40/40/42.
Caveats: tiny samples (n=3 per arm), one model family — the numbers shouldn't be quoted as established. But the pattern matters: companies selling agent-produced research have briefing chains (research → synthesis → client), and "our agents read the sources" is an attack surface if omission alone can swing conclusions. The author proposes an "integrity layer": provenance tracking, points-first formats, omission audits.
More from Safety
- 'Ten years hiding in one of the most read codebases': open source's many eyes theory under fire — aronchick · 2026-10-22
- METR red team: two outsiders found ways OpenAI's AI agents could slip past their monitoring system — ross2000 · 2026-10-22
- Composio launches Agent Firewall, an SLM-powered runtime boundary for agent actions — arankomatsuzaki · 2026-10-22
- METR and Redwood embedded at OpenAI for three weeks to assess internal agent misalignment monitoring — CFGeek · 2026-10-22
- Daniel Rock: HFT's market lessons — verification, circuit breakers and an 'immune system' for the internet — danielrock · 2026-10-22
- Security researchers: nation-states already deploying abliterated open-weight models for offense — joshua_saxe · 2026-10-22