Brief IA
›

HD Moore Tracks AI Leaks with Baited Secrets

Models & LLM·Tom Levy·

HD Moore Tracks AI Leaks with Baited Secrets

HD Moore Tracks AI Leaks with Baited Secrets
Key Takeaways
1HD Moore claims to have sent nearly two million requests containing traceable secrets.
2He reports request thefts at three hacked hosting providers.
3A secret URL was reportedly accessed more than two months after being sent to a U.S. hosting provider.
Why it matters : transport encryption does not prevent intermediaries from reading and then, according to their rules, retaining the requests.

HD Moore, Director of Research at Dragos, a security company, announced on Thursday that he has sent nearly two million requests to AI services with traceable secrets. He reports thefts at three compromised hosting providers, as well as the reuse of information despite chosen privacy settings.

URLs Accessed Despite Privacy Rules

A confidential address accessed more than two months after a single inference call: this is one of the behaviors HD Moore claims to have observed at a U.S. hosting provider. Another case involved a scraping tool returning to the received addresses several days after each request, regardless of the router used.

The case of a Chinese AI lab spans from July to October. According to HD Moore, its systems accessed the secret addresses contained in more than half of the requests sent to its models. This proportion reached about two-thirds for those accompanied by a directive prohibiting collection or a Zero Data Retention (ZDR) tag. The consultations typically occurred after two or three days.

Several U.S. hosting providers market the models from this lab. According to the researcher, requests passing through these providers triggered the same extractor and returned to the Chinese lab. He also reports that a client git, hosted in another cloud, subsequently attempted to copy the private repositories mentioned in the requests. These attempts exclusively concerned those using the lab's models.

Stolen AWS Keys at Hosting Providers and Resellers

At one of the hacked providers, a few hundred AWS keys were reused by the attackers, claims HD Moore. All experimental requests sent during brief windows at the end of October would have been stolen, including some that were denied before any model execution. For the researcher, this last observation places the copy at the entry point of the service, rather than in the internal processing layers.

In the other two compromises, requests marked as ZDR were among the stolen items, he reports. Individuals linked to the attacks tested the keys, sometimes after just a few hours, to determine their permissions. They searched for other identifiers in S3 spaces, secret managers, and parameter stores, attempted to activate models via AWS Bedrock, and added their own IAM users to maintain access. According to HD Moore, tens of thousands of calls to AWS programming interfaces were thus made using the stolen keys.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

From a single rented server, secrets transmitted to two compromised hosting providers were used together, almost immediately. These facts suggest that the two providers were almost certainly hacked by the same adversary.

Other reuses involve the resellers themselves: some attracted requests with low prices, then activated the registration. Identifiers from these registrations ended up being used by third parties. To isolate their responsibility from that of the model provider, HD Moore passed requests to the same model through different sellers and compared the observed reuses.

Intermediaries Can Read Every Request

The setup relied on previously unused secrets: keys and identifiers granting access to empty AWS accounts, passwords notably for SSH, unique web addresses, DNS names, and references to private git repositories. HD Moore controlled the infrastructure receiving the connections made with these lures. Each secret appeared only in a request sent to a single provider. These markers are nicknamed "canaries," in reference to the birds used in mines, because they serve here to identify the links responsible for a leak.

The targeted services are part of a chain where applications pass through intermediaries like OpenRouter or Vercel. Rather than directly soliciting OpenAI, Anthropic, or another major provider, a developer can delegate model selection to them. These routers direct requests based on factors such as price or the service's suitability for the task.

Encryption during transport does not prevent successive actors from reading the content: it is decrypted at each step. The router examines the text to select an inference service, then may direct it to a set of models operated at different hosting providers. A request can even pass from Vercel to OpenRouter. Each intermediary is likely to record its content and retain it according to its own rules, potentially without a deadline.

Moore Recommends Choosing Each Provider

To limit leaks, HD Moore favors models hosted on the servers of the user organization. This option remains costly. When self-hosting is financially unfeasible, he instead recommends a direct relationship with a major provider accompanied by a contract and a consultable privacy policy.

This direct relationship does not guarantee airtightness, the researcher notes, but he believes it reduces risks. He also advises explicitly selecting providers in advance, rather than retaining the sets of providers or resellers offered by default.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.