Relatório de inteligência

Two million bait secrets expose leaks across AI routers and inference resellers

A Dragos researcher seeded trackable credentials into AI prompts and traced them to breached providers, logging resellers and ignored privacy flags.

INSIGHTEX Editorial Desk

2 min de leitura

Two million bait secrets expose leaks across AI routers and inference resellersRelatório de inteligência

What happened

Security researcher HD Moore, chief research officer at Dragos, sent close to two million AI prompts through a network of inference providers and resellers, each carrying a unique secret he could trace, Ars Technica reported. The bait included working AWS keys linked to empty accounts, SSH passwords, one-off URLs and DNS names, and references to private git repositories. Each secret went into a single prompt sent to one provider.

Because each secret was used only once, Moore could tell which party had exposed it when it resurfaced. He found signs that attackers had breached at least three hosting providers and copied incoming requests. In one breach, a couple hundred of the bait AWS keys were reused. Across the experiment, the keys drew tens of thousands of AWS API calls, and intruders tried to enable models in AWS Bedrock and add their own IAM users. Some low-priced resellers also logged prompts, and the credentials in them were later used by others.

Moore also reported that providers disregarded Zero Data Retention and no-data-collection flags. A Chinese AI lab's systems fetched secret URLs from more than half the prompts sent to its models between July and October, including about two-thirds of those flagged for privacy. US hosts reselling the lab's models triggered the same scraper. One US provider crawled a secret URL more than two months after a single call.

Context

AI routers take prompts from app users and send them to whichever inference engine best fits criteria such as price or task, so developers do not have to contract directly with Anthropic or OpenAI. Prompts are encrypted in transit but decrypted at every hop. Moore said on-premises hosting is safest but costly. Otherwise, he advised working directly with a top-tier provider and choosing upstream providers instead of accepting default pools.

FonteArs Technica

O briefing

A semana em tecnologia, em um só e-mail.

As reportagens que valem o seu tempo, toda semana. Cancele quando quiser.