97.2% of Safety Blockers Stop Blocking After Multihop Agent Handoffs, arXiv Paper Finds
ombre-ai · reddit · 2026-10-30
The paper "When 'Must' Becomes 'Maybe'" (Sun et al., revised Oct 13) studies how safety constraints degrade during multi-agent handoffs, with permissions fixed and only the handoff artifact changing.
Key findings:
- Direct state passing preserves all blockers; compression, plan assimilation, or ownership deferral weakens them — 76.7% deactivated under ownership deferral, 97.2% under multihop compression, turning "must" into an acknowledgeable caveat.
- Restoring full blocker descriptions in fixed-artifact replays lifted recognition from 6.2% to 99.2% and cut forbidden actions from 85.2% to 0%.
- A prompt-level authorization check cut forbidden actions to 1.6%, yet the handoff itself was still 95.3% deactivated — final actions can look fine while the artifact is damaged.
The setup uses controlled synthetic data, so real-world failure rates are unproven. The authors suggest a simple test: ask your production agent which constraints are hard stops versus context after a handoff.
More from coding & agent
- fx: A 6 MiB Zig-Native Open-Source Coding Agent That Runs in the Browser — cramforce · 2026-10-30
- Ontographia: LLMs Emit Constrained Intent, Never Raw Cypher — JeremyCMorgan · 2026-10-30
- Coinbase for Agents goes live on NousResearch Hermes with persistent memory and MCP trading — MurrLincoln · 2026-10-30
- Indie dev launches one-off-purchase WYSIWYG editor Better Rich Text, itemizing TinyMCE and CKEditor fees — CtrlAltDwayne · 2026-10-30
- Studio owner builds AI voice receptionist in 12 minutes on Cartesia Managed Agents — aftahi_ai · 2026-10-30
- 63 days left in 2026: a 6-step plan to restart your career in applied AI — brandon_galang · 2026-10-30