2026.02/AI Safety
Manifold of Failure: Behavioral Attraction Basins in Language Models
S. Munshi, M. Bhatt, V. S. Narajala, I. Habler, A. Al-Kahfah, K. Huang, B. Gatto
Abstract
Maps unsafe behavioral regions in language models as attraction basins, using quality-diversity search to expose distinct vulnerability niches and their geometry.
Where it sits
Pillar 02 — Adversarial Measurement
Where does the system actually break?
Find where systems fail under attack — harder evaluations that expose weaknesses static benchmarks miss.
Outside writing
- The Weather Report ↗Independent article explaining the MAP-Elites vulnerability map and its results.
More from measurement
Whole pillar →2026.03LAAF: Logic-layer Automated Attack Framework — A Systematic Red-Teaming Methodology for LPCI Vulnerabilities in Agentic Large Language Model SystemsIntroduces an automated framework for testing logic-layer prompt-control injection across persistent memory, retrieval pipelines, tools, and multi-stage agent lifecycles.2026.01Large Empirical Case Study: Go-Explore adapted for AI Red Team TestingAdapts Go-Explore to agent red teaming and shows that random-seed variance can dominate algorithm choices, making multi-seed evaluation essential for defensible comparisons.2025.07Logic-layer Prompt Control Injection (LPCI): A Novel Security Vulnerability Class in Agentic SystemsDefines logic-layer prompt-control injection, where delayed or conditional payloads persist through memory, vector stores, and tool-connected agent systems.