Anthropic Warns Single-Turn Safety Tests Miss Multi-Turn AI Misuse
Emerging Technology Risk · 3 sources · Updated 4d ago · First seen Nov 17, 2026

Anthropic warned that increasingly capable AI models are outgrowing common safety benchmarks, allowing strong test scores to obscure weaknesses relevant to cybersecurity misuse, fraud and other harmful behavior. The company reported aggregate harmless-response rates of 94%–98% in single-turn evaluations on its API, but said multi-turn evaluations produced a wider range of results. Malicious users can gradually escalate requests across a conversation to obtain harmful information that models would refuse when asked directly. Anthropic also identified unrealistic scenarios, contaminated test data, reward hacking and unreliable automated graders as factors that can undermine safety conclusions.
Following unspecified evaluation incidents, Anthropic said it hardened its testing environments, temporarily stopped many internet-enabled evaluations and introduced real-time monitoring for web-search evaluations requiring internet access. It recommended realistic, sandboxed, multi-turn testing with difficulty calibrated to model capabilities, broader coverage and domain-specific scoring rather than reliance on a single aggregate safety score. For enterprise assurance, its guidance emphasizes human validation of automated graders, preregistered analysis, transparent reporting of uncertainty and deployment configurations, and continuous testing throughout model development. Anthropic also called for responsible disclosure of severe flaws in deployed systems.


