Conflux Prevents Privilege Escalation in LLM Agents
Original title: Conflux: Preventing Privilege Escalation with Principal Contexts and System-Level Evaluation
Conflux is a kernel-based defense using principal contexts and ITES to enforce authority from all influencing principals in LLM agents.
LLM agents combine information from multiple principals while acting externally. The paper formalises privilege escalation as a system-level objective where an action proceeds only if every influencing principal would allow it. Conflux presents a principal-context defence with kernel ITES that derives authority from provenance and intersects permissions of every influencing principal.
Unconfirmed: Evidence is bounded and partial; no unbounded deployment guarantee
- ITES intersects permissions of every influencing principal
- Conditional authority-confinement theorem proven under complete mediation
- SLED black-box evaluator exhaustively explores finite model
- SLED-V white-box evaluator checks serialisable representation
- Evidence is bounded and partial
Impact: Prevents privilege escalation in LLM agents
From the original article
1 min read at arxiv.orgLLM agents combine information from multiple principals while acting externally. We formalise privilege escalation (PE) as the system-level security objective: an action may proceed only if every principal that influenced it would allow it. Model-level defences remain probabilistic, so enforcement must sit outside model judgement.
By Raya Buckley · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.