OpenCode agent extracts 16-character system prompt flag from LLMs

arxiv.org1 min read

Researchers audited 363 text-in/text-out models from OpenRouter and found that 300 leaked a protected 16-character flag embedded in their system prompt. The study used an AI agent steered by a human researcher to persuade the models to disclose the secrets. Over half of the successful captures involved a non-English language pivot, most frequently French. Open article →