1 min read
Add as a preferred source on Google

Dragos Study Finds AI Prompt Logs Exposed to Sellers and Hackers

Dragos Chief Research Officer HD Moore published findings Nov. 19 after tracking unique fake links and passwords placed in about 2 million prompts from July through November.

A sealed security module rests beside blurred prompt pages / TokenPost.ai
A sealed security module rests beside blurred prompt pages / TokenPost.ai

Dragos Chief Research Officer HD Moore published a study Nov. 19 that found some inference marketplace sellers read customer prompts and a Chinese AI lab opened links in prompts marked for zero data retention.

The team planted unique fake links and passwords in about 2 million prompts from July through November, then tracked whether they were used. Some sellers on inference marketplaces enabled logging and read customer prompts. The activity was tied to individual sellers, not the marketplaces as a whole.

The Chinese AI lab opened links from about two-thirds of prompts carrying the zero-data-retention setting, including some prompts sent through U.S. resellers. The setting is meant to stop providers from storing prompts.

Hackers also had access to three hosting providers. The providers were unaware of the intrusions until Dragos notified them.

Several projects are developing ways to verify how AI prompts and models are handled. Antseed lets buyers require sealed hardware designed to prevent even the machine’s owner from reading prompts, and fingerprints sellers to catch swapped models. Anthropic has open-sourced a prototype intended to verify which model produced an answer. Pearl is developing a system with Attestable.

MEGA co-founder Shuyao Kong said the company is building verification and prompt scrubbing into its inference marketplace.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…