Code and data used in our NeurIPS 2026 paper: AI Agents May Always Fall for Prompt Injections
-
Updated
May 19, 2026 - Python
Code and data used in our NeurIPS 2026 paper: AI Agents May Always Fall for Prompt Injections
Leakage-verified contextual-integrity benchmark: does an agent disclose a sensitive attribute to a recipient the context forbids. Defensive, synthetic-only. Sibling of leakgauge.
Research notes and surveys on privacy-preserving memory for AI agents: machine unlearning, contextual integrity, differential privacy, and memory-architecture defenses.
Generative model of Contextual Integrity (CI) in multi-actor LLM systems: scenarios from CI taxonomies, generation pipeline, multi-agent simulation, and a two-sided appropriateness metric. No actor-level lever closes the protection/utility gap.
Need-to-Know (NTK): the open protocol for what an AI agent may say to other agents. HCP decides what your agent can know. NTK decides what it can say. Python stdlib, CLI, MCP server.
How much private information leaks incidentally between LLM agents as a function of communication topology — an evaluation built on Terrarium.
Mechanistic study of contextual-integrity post-training in Qwen2.5-7B, testing whether improved privacy behavior comes from new mechanisms or better use of machinery already present in the base model.
To associate your repository with the contextual-integrity topic, visit your repo's landing page and select "manage topics."