KillSwitch language challenges LLMs, Claude Opus scores highest
A developer created KillSwitch, an esoteric stack‑based programming language designed to be hard for large language models (LLMs) and coding agents. The language embeds hidden falsehoods, requires counting letters in identifiers, and includes phrases that trigger model safety guardrails. Most script text is ignored by the interpreter, and literals can be referenced by line numbers, making the…
Key points
- KillSwitch is a stack‑based esoteric language built to exploit LLM weaknesses such as letter counting and safety‑guard triggers.
- Claude Opus 5 topped the benchmark with a score of 66.8, while open‑weight MiMo‑V2.6‑Pro was the best non‑proprietary model.
- Fable 5.1 and Gemini 4 Argon performed poorly, reportedly due to overly sensitive safety guardrails.
The author benchmarked several major LLMs and coding agents on tasks such as implementing a hello‑world program, explaining script output, and modifying temperature conversion code. Claude Opus 5 achieved the top score of 66.8, outperforming other proprietary models. Open‑weight MiMo‑V2.6‑Pro was the best non‑proprietary model. Models like Fable 5.1 and Gemini 4 Argon performed poorly, reportedly because their safety guardrails are overly sensitive. The benchmark results are posted at bench.killswitch‑lang.org, and the author plans to expand the task set for future model releases.
Model pages: Gemini 4 Argon → · MiMo-V2.6-Pro →
The story so far
6 episodes →- KillSwitch language challenges LLMs, Claude Opus scores highestthis story
I built a programming language that LLMs can't use
blog.picheta.me · 17 October 2026
Loading the full article…
This text was published by blog.picheta.me and written by Dominik Picheta. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Hacker News discussion · 47 pointsnews.ycombinator.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- The Truth Behind 'Apple and Meta Team Up on AI': It's Not Siri, It's Mac Studio | A Manager's Breakdown of Confirmed Facts, Reports, Unannounced Details, and Im · 1 src
- OpenAI makes Prompt-to-Voice generally available for developers · 3 src
- Meta releases Muse Spark Open, a 816B-parameter multimodal model with 1M-token context · 37 src
- Anthropic’s Claude Opus 5.5 tops Arena AI model rankings · 2 src
- Modders report DLSS 5 yields up to 90% performance boost on unsupported games · 2 src
Comments
via GitHub Discussions