Article

Measuring What Survives: Pressure-Filtered Persistence and a Masking Index for LLM-Agent Safety Evaluation

2026-06-16

Abstract excerpt

<title>Abstract</title> <p>When a safety intervention lowers an agent’s attack-success rate, two different things may have happened: the intervention removed a load-bearing failure route, or the agent learned the measured attacks while the route survived. Frequency metrics are necessary, but they do not by themselves identify which behavioural structures persist as adversarial pressure rises. We propose a measure...

Topics

Open a Topic to create a Post that cites this publication.

Identifiers and source

Literature Corpus work: dbacc19d-3ed5-5aeb-a932-c2798d36df34
DOI: 10.21203/rs.3.rs-10026792/v1

Open publication

Related research

Semantic proximity does not establish scientific evidence.

Click a neighbor to travelStep 1 · 12 closest

Measuring What Survives: Pressure-Filtered Persistence and a Masking Index for LLM-Agent Safety EvaluationDOI 10.21203/rs.3.rs-10026792/v1Read

Protocol-Dependent Resolution in Finite-Sample Model SelectionSemantic proximity Read

The World Is a Hypothesis: Epistemic Control for Executable World-Model Agents in Novel Interactive EnvironmentsSemantic proximity Read

Multi-Agent Human-AI Systems with Low-Code Platforms Enabling Adaptive Web Services and Real-Time Anomaly Remediation in Distributed ArchitecturesSemantic proximity Read

Agentic AI and LLM-Driven Framework for Contextual Anomaly DetectionSemantic proximity Read

A Layered Guardrail Architecture for Agentic AI in Cloud-Native DevOps Pipelines: Design, Implementation, and Empirical Evaluation on Live AWS InfrastructureSemantic proximity Read

Naive Defect-Recidivism Mining Is Inflated by Agent Workflow Artefacts: A Construct-Validity Study at Corpus ScaleSemantic proximity Read

Structural Leakage in Host Intrusion Alert Corpora: An Evaluation Framework for ATT&CK Technique and Tactic MappingSemantic proximity Read

Who Broke the Pipeline? Traceability andAccountability in Multi-Agent LLM SoftwareEngineering PipelinesSemantic proximity Read

Mutation Testing of Task-Scoped State Oracles in Software-Agent Benchmarks: A Cross-Benchmark Empirical StudySemantic proximity Read

The Invariant Energetic Cost of Constraint Persistence: A Self-Referential Framework for Measuring Alignment RobustnessSemantic proximity Read

Detection and Mitigation of Mythos-Class Frontier Model Capabilities: A Layered Reference ArchitectureSemantic proximity Read

Limits of Self-Correction in LLMs: An Information-Theoretic Analysis of Correlated ErrorsSemantic proximity Read