2 articles

#scalable-oversight

$A mesh of twelve AI agent nodes rendered in brushed aluminum, three showing amber failure signatures with fractured surfaces and severed fiber-optic strands, the remaining nodes connected by cyan light conduits against a dark navy background$

Researchreport

The Coordination Failure Catalog: How Multi-Agent Systems Break in 2026

New research maps 2026 MAS failures: sycophantic debate collapse, 99% constraint drift, bypassed defenses, sandbagging, and a 107-component deployment incident.

May 6, 202613 min read

Researchbreaking

Anthropic/MATS/Redwood: Weak Models Can Correct AI Sandbagging

Anthropic/MATS/Redwood paper: weak-supervisor training stops capable AI sandbagging on tasks humans can't evaluate — scalable oversight milestone with direct AI safety implications.

May 6, 20261 min read

AI Intelligence Newsletter

Curated AI insights — sent when there's something worth your inbox.