KillSwitch: An Esoteric Language Designed to Deceive LLMs

Developer dom96 introduced KillSwitch, an esoteric language specifically created to mislead LLMs, and the KillSwitch-Bench benchmark, where Claude Opus 5 leads on adversarial tasks with 66.8% at $1.49 per task, while the best open-weight model is Xiaomi MiMo-V2.6-Pro with 38.3%.

Compiled by Sergey KostenchukPublished 2026-10-18Updated 2026-10-18

2026-10-18 Research Anthropic

🤖 KillSwitch: An Esoteric Trap Language for LLMs

Developer dom96 released the esoteric language KillSwitch: its constructs trigger model safety filters, the number of letters in identifiers changes code behavior, and SPEC.md hides lies invisible to humans. In KillSwitch-Bench, coding agents solve simple tasks in containers without internet (limit of $5 and 30 minutes, programmatic evaluation). The leader is Claude Opus 5 (66.8%), and the best open-weight model is Xiaomi MiMo-V2.6-Pro (38.3%).

🌍 The benchmark measures not the ability to write code, but resilience to adversarial specifications and false hints—this is how agents most often break in production. A typical failure is Failed verification: the code works, but does not match the language logic.

👤 Everything is open: a table of 20 models with prices and times is available at bench.killswitch-lang.org, and the interpreter and SPEC.md are on github.com/dom96/killswitch. Source 1: https://bench.killswitch-lang.org/ Source 2: https://news.ycombinator.com/item?id=50139549

Sources