Dragos finds evidence of breaches at three superintelligence hosts

Researcher HD Moore says attackers used some credentials planted in an experiment involving nearly two million prompts, while some requests marked for privacy were also exposed. The findings point to risks as prompts pass through multiple providers.
In this article
Security firm Dragos found evidence that attackers compromised at least three superintelligence (SI) hosting providers and used credentials planted in test prompts, its chief research officer HD Moore says. The experiment shows how sensitive requests can pass through several companies before reaching a model.
Moore published the findings on 19 November, after the first detection in mid-July and another set of intrusions in late October. Dragos began notifying providers about separate data-sharing issues in early November.
Moore, who created the security testing framework Metasploit, placed unique tracking items known as canaries in each request. They included working AWS access keys tied to empty accounts with full logging. When a key or link was used, the team could trace it back to the original prompt, route and privacy setting.
What the planted keys revealed
In the largest alleged breach, Moore says, a couple hundred test keys were taken from requests sent to one provider during short windows in late October. Some requests were rejected before a model ran, which he says indicates that they were copied as they arrived at the provider.
The bait keys were used in tens of thousands of AWS API calls, Moore says. Those using them searched for data, tried to enable hosted models and attempted to create additional access. Keys sent to two providers were used together from one rented server, suggesting to Moore that the same attacker had compromised both.
Dragos directly observed attempts to use its own credentials against controlled systems. Moore says real customer requests were processed alongside the bait and were almost certainly copied too. Dragos has not named the three providers, and Ransomnews says it could not independently verify the tests because the providers were unnamed.
How prompts pass between companies
A gateway routes requests to available models, sometimes through other gateways and hosting providers. Moore says each company in the chain can see the prompt and response; a request through Vercel, for example, can pass through OpenRouter before reaching another provider.
Moore says marketplace resellers could retain prompts using ordinary request logging. In one comparison, someone opened test links sent through one seller, while prompts sent to the same model through another seller were untouched. That pointed to the seller rather than the model alone. Because a canary does not by itself identify the leaking company, Dragos also traced routes using requests and the signatures in their responses.
Privacy settings and delayed visits
Zero Data Retention is a setting intended to prevent a provider from storing a prompt and its response. Moore says some test requests stolen in the other two alleged breaches carried that flag.
Separately, Moore says an unnamed Chinese SI lab’s systems fetched secret links from more than half of the prompts sent to its models between July and October, including roughly two-thirds of those marked Zero Data Retention or “no data collection.” The visits typically came two or three days later. Requests sent through several US hosts reselling the lab’s models triggered the same automated system.
Moore says one US provider continued to crawl a test link more than two months after a single request, while another returned to links days after requests.
Provider responses and precautions
Moore says one of the three providers published a preliminary security notice, several providers began investigations, and a gateway told Dragos it had stopped forwarding traffic to the marketplace. Dragos is working with about a dozen providers and Moore expects that number could grow.
For sensitive work, Moore recommends running a model on an organization’s own equipment if it can afford to. Otherwise, he suggests contracting directly with a provider, while warning that a breach or data-sharing agreement can still expose prompts. Teams using gateways should choose the upstream providers themselves and check that privacy settings affect where requests go, he says.
He also recommends planting test credentials in the information sent by automated agents and treating their use as an incident. Filters can remove known credential formats, he says, but will not catch every kind of sensitive information.
Follow AveNews
Earlier in Technology
Chips & ComputingMexican mobile users report lost signal and failed calls
Telcel says service was intermittent in some localities and users were reconnecting, while AT&T Mexico says its network was operating normally.
2 min
SuperintelligenceUS House committees urge limits on Chinese superintelligence models
Their investigation identifies uses at Airbnb, DoorDash and Cursor maker Anysphere, and recommends restrictions for sensitive systems; the proposals are not adopted rules.
4 min
SoftwareWhatsApp patches Android flaw that could expose users’ IP information
WhatsApp’s security team found the flaw while disrupting a scam campaign. The company suspects a scammer used it to make a message seem more credible and has fixed two other security flaws.
3 min
