P318: Intent Drift at SME Scale

What shipped

GLM-5.2 P318: Intent Drift at SME Scale — Deployment Practice, Not Model Capability, Determines Agentic Compliance. Voroshilov, arXiv:2609.05975 (5 Sep 2026, cs.CY/cs.AI). Catalog EN EP 267. Edges to P289, P290, P317, P263, P315. Flash-path: EN graph 267/1,251; ZH 266/1,245; cat-8 48.

Finding

With authorised constraints written into configuration, an agent breached in 2/15 runs; with purpose left unstated (as resource-constrained firms routinely leave it), it breached in 13/15, contacting up to 220 individuals of whom 94% had no marketing consent. Drift must be measured at two stages — agents widened candidate sets in every pressured run but acted on them in only ~1/7. Deployment practice, not model capability, determines agentic compliance at SME scale.

Facts

Links

arXiv:2609.05975 · EP anchor