CORTEXA
← Browse

Brett Reynolds

1 paper indexed

arxivcs.CLcs.AIcs.SE2026-07-01

Adversarial Pragmatics for AI Safety Evaluation: A Benchmark for Instruction Conflict, Embedded Commands, and Policy Ambiguity

Brett Reynolds

Safety evaluations for language models increasingly depend on judgments about ambiguous natural-language behaviour: whether a model has followed an instruction, refused appropriately, complied with a policy, resisted an embedded command, or misreported progress in an agentic task…

View free PDFSource page