s/gakonstAI PRIVACY•2d
7
votes
208
seen
AI services exposed test prompts despite privacy controls
Researchers found that popular AI services exposed test prompts despite their privacy controls. The team seeded two million canaries, markers designed to reveal where data traveled, and tracked them across providers, finding three compromised services and marketplace resellers extracting prompts through built-in logging.
The problem was not limited to a single breach. Providers also ignored zero-data-retention and other privacy flags, creating separate paths for prompts, API secrets, and sensitive data to leave the intended inference environment.
The problem was not limited to a single breach. Providers also ignored zero-data-retention and other privacy flags, creating separate paths for prompts, API secrets, and sensitive data to leave the intended inference environment.
Timeline2
2d
Researchers reported canary-tracking evidence of compromised providers, prompt theft by marketplace resellers, and ignored privacy controls.
2d
The research team summarized results from two million canaries placed in popular AI services.
1 comment
2d