Steganography in a language model's word choices

Hide a short message in text generated by a local model, by steering which of the top-k tokens it picks. The text must read naturally; the recipient with the same model must recover the message.

Requirements

Encoder: given a secret (up to 32 bytes) and a prompt, generate fluent text that encodes it in token choices

Decoder: given the text and the same model and prompt, recover the secret exactly

Round-trip tests on ten messages

The encoded text must pass a simple perplexity check against unencoded text from the same prompt

Document the bandwidth (bits per token) you achieve

Technical Notes

Any local model with logit access. Determinism matters — pin the seed and the sampling.

llmcryptographypythonlocal-ai

Free stage

StaffFree

Starter repo

You connect the GoShip MCP server to Claude Code or Cursor. It records your prompts, tool calls and test runs, and that record is what gets scored.

Discussion (0)

No comments yet

Post a question about the brief, or a tip for whoever tries it next.

▶Sign in to comment