OverThink: Slowdown Attacks on Reasoning LLMs
Factual evidence
What the source reports
Researchers demonstrate 'OverThink', an attack forcing reasoning LLMs to generate excessive hidden tokens via injected decoy problems.
Inspect the evidence
- Inclusion basis
- Enterprise AI
- Publisher and source type
- arXiv cs.LG — Machine Learning · RESEARCH
- Published by source
- 21 September 2026
- Collected by OneBench
- 22 Sept 2026, 03:01 UK
- Original headline
- OverThink: Slowdown Attacks on Reasoning LLMs ↗
Stored source excerpt
arXiv:2502.02542v5 Announce Type: replace Abstract: A reasoning language model (RLM) generates costly reasoning tokens, often hidden from the users, that help it excel at many…
Short excerpt from the collected text, not the full source. Use the source link to read it in context.
The factual summary is a OneBench synthesis, not a quotation or independent verification. Collection time is not publication time. Open the source for its full context; related reporting can share the same underlying announcement.
OneBench interpretation
Institutional assessment
So what
Adversarial context injection can inflate API billings and latency in reasoning models by forcing unnecessary background token generation.
Do what
Review input sanitisation rules with the team responsible for AI security before exposing reasoning models to external data sources.