Reflections on Trusting Trust, revisited: poisoning self-modifying AI coding

sbulaev · hn · 2026-09-18

An arXiv paper revisits Ken Thompson's classic "Reflections on Trusting Trust" in the age of AI coding: as coding tools start self-modifying and generating code, attackers could poison training data or generation processes, covertly spreading malicious code through self-modification loops. The paper analyzes these attack vectors and threats to the AI coding trust chain.

Original post →

More from Safety

Safety channel →