AI & softwareResearch findings
Covert Threat of Hiding Personal Attributes in LLM Outputs
Original published 2026.11.13 03:51 KSTAdded to Blackdeck 2026.11.15 02:57 KST
What has changed?
Sensitive user information is inferred from conversations and stealthily encoded into harmless-looking model outputs
Key points
- User personal attributes are inferred from routine conversations with LLMs
- Inferred data is encoded using invisible characters or subtle word choices
- Shared model outputs can serve as a covert channel for sensitive information leakage
Connect it to your work
Users unknowingly transmit personal data through shared model-generated content
What remains unproven
This research note is based on an abstract; check the original for test conditions and application scope.
Originals used for this note
Links open the original publisher in a new tab.