Poisoned Identifiers Survive LLM Deobfuscation: A Case Study on Claude Opus 4.6

TL;DR

This study investigates whether poisoned identifier names in JavaScript survive deobfuscation by large language models, revealing persistent patterns despite semantic correctness and task framing.

Contribution

It demonstrates that poisoned identifiers can persist through deobfuscation even when models understand the code semantically, highlighting challenges in model robustness.

Findings

Poisoned names persisted in every baseline run across artifacts.

Persistence occurred alongside correct semantic commentary.

Reframing tasks reduced propagation of poisoned identifiers.

Abstract

When an LLM deobfuscates JavaScript, can poisoned identifier names in the string table survive into the model's reconstructed code, even when the model demonstrably understands the correct semantics? Using Claude Opus 4.6 across 192 inference runs on two code archetypes (force-directed graph simulation, A* pathfinding; 50 conditions, N=3-6), we found three consistent patterns: (1) Poisoned names persisted in every baseline run on both artifacts (physics: 8/8; pathfinding: 5/5). Matched controls showed this extends to terms with zero semantic fit when the string table does not form a coherent alternative domain. (2) Persistence coexisted with correct semantic commentary: in 15/17 runs the model wrote wrong variable names while correctly describing the actual operation in comments. (3) Task framing changed persistence: explicit verification prompts had no effect (12/12 across 4 variants), but reframing from "deobfuscate this" to "write a fresh implementation" reduced propagation from 100% to 0-20% on physics and to 0% on pathfinding, while preserving the checked algorithmic structure. Matched-control experiments showed zero-fit terms persist at the same rate when the replacement table lacks a coherent alternative-domain signal. Per-term variation in earlier domain-gradient experiments is confounded with domain-level coherence and recoverability. These observations are from two archetypes on one model family (Opus 4.6 primary; Haiku 4.5 spot-check). Broader generalization is needed