← News

AI securitySources: arXiv · 16 Nov 2026

Symbolic Swapper attack degrades LLM planning by over 70%

An arXiv paper introduces an adversarial attack that renames the symbols in planning tasks, cutting success rates for base models and robust planning frameworks by more than 70%.

In plain words

AI agents that plan tasks often rely on readable labels, like action and object names, to understand what to do. The Symbolic Swapper tool automatically rewrites those labels into unfamiliar forms. The authors report this drops planning success by over 70% for some models, showing the models lean on those labels rather than true planning skill.

What was announced

Researchers behind the arXiv preprint 2611.17910 present an LLM-based attack model called Symbolic Swapper. It automatically generates obfuscated planning domains with non-standard symbolic vocabularies, meaning the names of actions, predicates, and objects no longer carry the commonsense cues that language models usually exploit. The authors state the goal is to test whether recent reasoning models, which improved on obfuscated planning tasks, can still plan when the vocabulary is systematically scrambled.

In their results, the attacks decreased planning performance of base models and even robust planning frameworks by over 70%. The authors also analyzed why performance broke down and found that successful obfuscation made models appear to significantly degrade in reasoning quality, even though the models showed some inherent ability to understand the domain. They describe this as a gap between perceiving a task domain and being able to plan under it, which they argue makes LLM agents less reliable than symbolic planners when commonsense cues are unavailable.

What it does not show

  • The evidence is limited to the paper's abstract, so specific models tested, benchmark details, and evaluation methods are not described.
  • The over-70% degradation figure is the authors' own reported result and has not been independently confirmed.

For security and IT teams

Scope
The findings come from a single arXiv abstract covering an adversarial attack framework for LLM planning agents; no deployment or mitigation guidance is included in the available evidence.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.