Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
Steganography Without Modification: Hidden Communication via LLM Seeds
TL;DR
Existing Large Language Models (LLMs) have an unexploited vulnerability that allows for hidden communication without altering their architecture. By leveraging the deterministic nature of pseudo-random number generators (PRNGs) in token sampling, a secret message can be encoded within the generated text.
✦ Why It Matters
Engineers and researchers can leverage this method for secure communication without altering existing LLM architectures.
Key Takeaways
- Explore using LLMs for secure message transmission.
- Implement the known-prompt setting for optimal accuracy.
- Consider the implications of prompt knowledge on security.
- Test recovery accuracy with varying token lengths.
How It Works
The method exploits the deterministic nature of LLMs, where the output is influenced by the PRNG seed. By encoding a secret message in this seed, the sender can generate text that contains hidden information.
The receiver, knowing the structure of the output, can reconstruct the probability intervals of tokens to deduce the original seed and thus the hidden message.