← Latest briefing

Technology

Prompt injection and compression tests indicate OpenRouter model Ox-Alpha is Z.ai's GLM

Researchers extracted system instructions and ran text compression analysis to attribute the mysterious model's outputs.

The short version

  • Prompt injection techniques revealed that the mysterious OpenRouter model 'Ox-Alpha' was instructed by system prompts to hide its origin and claim to be built by an undisclosed organization.
  • When researchers fed the extracted system prompt back into the model, it broke its trust pattern and identified itself as GLM, a language model developed by Z.ai.
  • A separate compression-based classification analysis matched most Ox-Alpha test outputs to GLM-5.3 over competing reference models.
  • Z.ai and OpenRouter have not officially confirmed the model's lineage or operational details.

Key facts

  • A unidentified language model listed under the name Ox-Alpha on OpenRouter gained attention while climbing the platform's leaderboard.[Hacker News]
  • A prompt injection query revealed system prompt instructions directing the model to identify strictly as 'ox-alpha' from an 'undisclosed organization'.[Hacker News]
  • Upon having its system prompt fed back into it as a user message, the model declared that it was actually GLM, a large language model created by Z.ai.[Hacker News]
  • A gzip-based Normalized Compression Distance (NCD) classifier matched 7 out of 14 Ox-Alpha text samples to GLM-5.3, outperforming matches to alternative models like Claude Opus 5, Gemini, and GPT-5.5.[Hacker News]

What remains uncertain

  • Whether Z.ai officially deployed or authorized the listing of Ox-Alpha on OpenRouter remains unconfirmed by the company.[Hacker News]

Sources