Salesforce proposes ErosionBench to evaluate multi-agent alignment decay

1. 17:40
   arXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesSelected

   ErosionBench: Probing alignment decay in enterprise multi-agent systems

   Salesforce AI Research proposes ErosionBench, a safety benchmark that evaluates cross-session alignment decay in enterprise multi-agent systems within an executable environment. Attacks propagate across sessions through poisoned shared state, long documents and trusted peer messages; no single message contains malicious content, yet unsafe actions assemble in the final tool call. The benchmark covers four decay mechanisms across four domains—quantitative finance, CI/CD supply chain, data engineering and autonomous CRM—for a total of 1,728 scenarios (1,296 malicious, 432 benign controls), annotated according to the OWASP Top-10, and additionally includes 198 image attacks each paired one-to-one with a text version.

Current attention 39·Peak in the comparable period 39 (Nov 10 22:00)·Change over the comparable past 24 hours 0%