NeFut Logo NeFut
Admin Login

[CS.AI] Analyzing the Analyzer: Self-Validating LLM Hazard Analysis with Constitutional Meta-STPA

Published at: 2026-07-11 22:00 Last updated: 2026-07-13 08:40
#algorithm #AI #Open Source

Abstract

Large language models (LLMs) are increasingly trusted to draft artifacts of safety analysis such as losses, hazards, Unsafe Control Actions (UCAs), and safety constraints within rigorous processes like Systems-Theoretic Process Analysis (STPA). However, a blind spot exists in this rapidly growing literature: every system is analyzed except the LLM-assisted tool itself, which is a safety-relevant system that can hallucinate standards, emit unverifiable constraints, and leave no audit trail from prompt to artifact. We take seriously the question that the field has overlooked — "who analyzes the analyzer?" and answer it by applying STPA to the tool itself. We present Constitutional Meta-STPA, an LLM-assisted STPA tool built around a closed loop: the tool conducts a meta-STPA of the class of AI-assisted safety tools and derives its governance constitution from the resulting loss$\to$hazard$\to$UCA$\to$constraint chain, yielding a published constitution of 21 Tool Principles and 8 Meta-Safety Principles, each bound to a code enforcement point.

We formalize the measured object as a constitution-marginal coverage operator over a principle set $P$ ($|P|=29$) with a soundness lemma that isolates coverage from model and scanner, and report four findings:

  1. Self-derivation: a frontier ensemble (claude-opus-4.8${+}$claude-sonnet-4) recovers 18/21 canonical and all 8/8 governance principles from the tool's own design, while a weaker pair recovers 12/21 and 3/8, indicating that the meta layer is model-limited, not constitution-limited, and the same 8/8 re-emerge from a second, independently authored tool.

Blogger's Review: This study delves into the self-analysis mechanisms of LLM tools, revealing potential risks that must be considered in the design of safety analysis tools. By introducing meta-STPA, the authors advance the security research of LLMs and provide a crucial framework for the governance of future AI tools. The emphasis on the tool's self-derivation capability showcases how to ensure the effectiveness of safety principles in complex systems.

Original Source: https://arxiv.org/abs/2607.08054

[h] Back to Home