Attack Success Rate Is Not a Number: On Measurement Validity in Agentic AI Security Evaluation

Abstract

Attack success rate (ASR) is the headline metric in nearly every published evaluation of attacks on, and defenses for, LLM agents. We argue that ASR as currently used is not a single quantity but a family of metrics parameterized by six design choices that papers seldom specify and never hold constant across the literature. We support this with two studies that require no proprietary access. First, a full-text meta-analysis of 259 agentic-security papers posted to arXiv between February 2025 and September 2026 finds that most report neither a variance estimate nor repeated runs for their headline attack metric: 58% (95% CI 44-71) in a hand-coded random sample of 50, 65.3% by automated coding of all 259. Only 30.9% disclose enough about decoding to establish whether their evaluation was even stochastic, and of the 64 papers we confirm use an LLM judge, 29.7% report any agreement check against human labels. Second, an analytical study shows that these omissions are not cosmetic: on a 100-instance benchmark, the minimum difference in ASR detectable at conventional power is 18.2 percentage points, and two defenses whose true ASRs differ by 5 points are ranked in the wrong order by a single-run evaluation roughly 21% of the time. Because several of the six axes shift ASR in a system-dependent way, the resulting incomparability is not a constant offset that cancels in comparison. We conclude that cross-paper ASR comparison is currently unsupported, and propose a ten-item reporting checklist targeted at each failure we measure. Our aim is not to dispute any individual result but to supply the shared measurement contract the field has so far done without.