# Adversarial Pragmatics for AI Safety Evaluation: A Diagnostic Framework and Seed Benchmark for Language-Mediated Control

arXiv: https://arxiv.org/abs/2607.01153 · HTML: https://mlsanity.com/paper/2607.01153/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Placed in Agent threats, guardrails and governance, one of 16 directions in LLM agents and tool use.
- Builds on 21 of its 41 references that are in the map.
- First posted on arXiv 1 Jul 2026.

## Placed in

- LLM agents and tool use / [Agent threats, guardrails and governance](https://mlsanity.com/agents/agent-threats-guardrails-and-governance/)

## Builds on (most cited first, 21 of 21)

- [Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena](https://mlsanity.com/paper/2306.05685/) (2023)
- [G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment](https://mlsanity.com/paper/2303.16634/) (2023)
- [Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection](https://mlsanity.com/paper/2302.12173/) (2023)
- [HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal](https://mlsanity.com/paper/2402.04249/) (2024)
- [Instruction-Following Evaluation for Large Language Models](https://mlsanity.com/paper/2311.07911/) (2023)
- [Refusal in Language Models Is Mediated by a Single Direction](https://mlsanity.com/paper/2406.11717/) (2024)
- [XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models](https://mlsanity.com/paper/2308.01263/) (2023)
- [The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions](https://mlsanity.com/paper/2404.13208/) (2024)
- [Evaluating Large Language Models at Evaluating Instruction Following](https://mlsanity.com/paper/2310.07641/) (2023)
- [Frontier Models are Capable of In-context Scheming](https://mlsanity.com/paper/2412.04984/) (2024)
- [Benchmarking and Defending Against Indirect Prompt Injection Attacks on Large Language Models](https://mlsanity.com/paper/2312.14197/) (2023)
- [AI Control: Improving Safety Despite Intentional Subversion](https://mlsanity.com/paper/2312.06942/) (2023)
- [Do-Not-Answer: A Dataset for Evaluating Safeguards in LLMs](https://mlsanity.com/paper/2308.13387/) (2023)
- [Tensor Trust: Interpretable Prompt Injection Attacks from an Online Game](https://mlsanity.com/paper/2311.01011/) (2023)
- [BetterBench: Assessing AI Benchmarks, Uncovering Issues, and Establishing Best Practices](https://mlsanity.com/paper/2411.12990/) (2024)
- [Can LLMs Separate Instructions From Data? And What Do We Even Mean By That?](https://mlsanity.com/paper/2403.06833/) (2024)
- [External Validity: From Do-Calculus to Transportability Across Populations](https://mlsanity.com/paper/1503.01603/) (2015)
- [IHEval: Evaluating Language Models on Following the Instruction Hierarchy](https://mlsanity.com/paper/2502.08745/) (2025)
- [Benchmark Inflation: Revealing LLM Performance Gaps Using Retro-Holdouts](https://mlsanity.com/paper/2410.09247/) (2024)
- [When LLM Judge Scores Look Good but Best-of-N Decisions Fail](https://mlsanity.com/paper/2603.12520/) (2026)
- [Safety is Contextual, LLM-Judges Are Not: Navigating the Rigid Priors of Evaluators](https://mlsanity.com/paper/2606.07874/) (2026)

## Built on by (most cited first, 1 of 1)

- [When benchmark inferences do not compose: Projectibility in AI evaluation](https://mlsanity.com/paper/2607.26159/) (2026)

## Abstract (arXiv)

Safety evaluations for language models increasingly depend on judgments about ambiguous natural-language behaviour: whether a model followed an instruction, refused appropriately, complied with a policy, or misreported progress in an agentic task. Existing benchmarks compress these into pass/fail labels, obscuring whether failures reflect capability limits, policy ambiguity, instruction conflict, scaffold failure, or unstable evaluator judgments. Adversarial pragmatics is safety-relevant model behaviour under instruction conflict, embedded commands, quotation, scope ambiguity, deixis, and indirect speech acts. It's designed to extend to multi-turn agent transcripts, but the seed set represents that family with a single-turn tool-result contrast. This paper introduces a diagnostic framework, an 18-item seed benchmark, a 54-row pilot, and a six-cell LLM-judge assessment, with a protocol keeping task success, policy compliance, risk, refusal, attribution, and confidence analytically separate. The benchmark separates four inference targets a single label can conceal: the regime-relative reference, configured-system behaviour, evaluator-output interpretation, and taxonomic assignment. Its intended use is diagnosis, not deployment certification, vendor ranking, or a general safety score. A first LLM judge that graded its own outputs with the expected answer visible missed the safety-relevant minority classes. Item-clustered intervals leave four of six chance-corrected agreement statistics unable to rule out a constant labeller, and hierarchical pooling shrinks the one eye-catching rubric effect toward the group mean and widens its interval through zero. Rejudging the objects across three judge models and two information conditions leaves the pattern intact: no cell recovers more than two of eleven partial successes, and the strongest cell's edge comes partly from never using that label.