ErosionBench: Probing Alignment Decay in Enterprise Multi-Agent Systems

Summary

ErosionBench is a benchmark that tests whether multi-agent LLM systems drift into unsafe tool actions as benign requests propagate through poisoned shared state, long-context documents and trusted peer messages. It covers 1,728 scenarios across four enterprise domains, plus 198 image-based attacks paired with text equivalents. A proposed Meta-Auditor reaches 98.5% recall at a 2.9% false-positive rate on a held-out split, and its detection is channel-invariant across image and text attacks.

Mitigation

The source proposes a calibrated Meta-Auditor pairing synchronous state checks with asynchronous trajectory monitoring as a reference defense.

Related items

• InfoAnthropic sets price range of $70 to $78 in upcoming IPO as AI lab marches toward NasdaqSame vendor · CNBC Technology
• InfoChatGPT’s Teen Mode could deprive teenagers of a fundamental need | Robin KirmanSame vendor · The Guardian Technology
• InfoAnthropic sharpens AI coding and security capabilities with Mythos 5.2Same vendor · CSO Online
• InfoClosing the AI guidance gap in educationSame vendor · OpenAI Blog
• InfoAnthropic's Claude Fable 5.2 Beats GPT-6 Cyber in Multi-Stage Cyberattack BenchmarkSame vendor · The Hacker News