Key facts
- Each secret went into one prompt, and each prompt went to just one provider.
- Moore controlled the infrastructure that would receive any request made with the credentials, so a triggered canary proved the data escaped and was acted on.
- The test targeted AI routers or gateways such as OpenRouter and Vercel that route prompts to inference engines by criteria such as the lowest price.
- The secrets were previously unused and included AWS access keys, SSH passwords, URLs, DNS names, and private git references.
Dragos researcher HD Moore ran a study from July through November 2026, seeding nearly two million one-of-a-kind tracking credentials and web addresses into AI inference requests. Each secret went into a single prompt bound for just one provider, traveling through services such as OpenRouter and Vercel so the source of any leak could be identified.
Moore seeded only unused material, such as AWS access keys tied to empty accounts, SSH passwords, web addresses, DNS names and private git references. The credentials pointed to infrastructure controlled by Moore, and because he received any request made with them, a triggered canary showed that data had escaped and someone had acted on it. He said the secrets would show which link in the chain was responsible for a leak. Crypto Briefing, describing HD Moore’s method, put it this way: “A canary that fires means the data escaped and someone acted on it, not that it theoretically could have.”
The researcher’s test focused on AI routers and gateways like OpenRouter and Vercel, which send prompts to inference engines based on factors such as the lowest price. Developers sometimes use an AI router to pick the most suitable model rather than signing with Anthropic, OpenAI or another large provider directly. Users of these routers may not see where their prompts end up, and Moore traced requests from people in the US to Chinese AI organizations.
Attackers compromised AI hosting providers in at least three cases and took customer prompt data. Targeted attacks during late October 2026 pulled about 200 unique AWS credentials, while confirmed theft activity stretched back to mid-July 2026. Attackers used some keys within hours of the requests, and the credentials produced tens of thousands of AWS API calls across services including S3 storage. Moore assessed that a provider blocking stolen requests before launch indicated intruders had copied data at the entry point, and intruders added their own IAM (identity and access management) accounts to keep access while seeking to activate AWS Bedrock models (Bedrock is AWS’s generative AI service) and check AWS Secrets Manager and Systems Manager Parameter Store.
Moore logged requests from a Chinese AI laboratory to secret URLs hidden in prompts, and they turned up in more than half of the laboratory’s model requests spanning July to October 2026. About two-thirds of those requests carried no data collection or Zero Data Retention labels. A seeded URL appeared in an internal Chinese company chat 24 days after the original request. Some US hosting providers resold the laboratory’s models, and requests routed through them triggered the same secret URL requests.
AI resellers and marketplaces captured customer prompts through built-in logging features, running counter to zero-data-retention policies. Under those policies, a provider processes a request and then forgets it. Moore said a zero-data-retention promise by itself is not enough, since a policy is only as strong as the weakest party handling the data. He said reseller logging can render a zero-data-retention guarantee decorative.
Providers affected by the breaches learned about them only when Moore notified them in November 2026. A public summary of his research appeared on November 19, 2026. Moore called the detection gap the biggest worry. He said intruders took data from at least three hosting providers over months, undetected until an outside researcher spoke up.
Moore advises organizations to self-host models where possible or to deal directly with a major provider rather than relying on standard pools or AI-router intermediaries. He said organizations should treat prompt contents as exposed, with customer records and internal links among the examples, along with cloud credentials and API keys.

Fireworks AI operates as a cloud platform that lets developers use open AI models and lets companies run and fine-tune them without handling infrastructure themselves. On Oct. 9, 2026, Fireworks AI told Uber that an unauthorized party had accessed Uber’s account using compromised credentials. Attackers reached Uber’s data on Oct. 7, 2026, two days before Uber received notice.
Uber cut data connections with Fireworks AI on Oct. 9, 2026, and finished its investigation on Oct. 13, 2026. Uber said the provider’s stolen credentials did not let attackers reach other Uber systems, keeping the breach inside the supplier’s environment. Uber’s Oregon filing counted 149,346 affected individuals, and the company said nearly 150,000 people had their data exposed.
Exposed information included names, addresses and birth dates. It also covered driver’s license numbers, license images and other government-issued ID details. Fireworks AI said in a November 9 update that attackers had accessed customer billing details and other secrets, along with datasets and unfinished AI models belonging to a small number of customers. Fireworks AI brought in Palo Alto Networks Unit 42 to investigate and notified law enforcement.
Fireworks AI said a financially motivated group tried to sell stolen AI access tokens through unauthorized services, a practice known as LLMjacking. No known cybercriminal group has taken public credit for the Uber data theft. Notion and Cursor said attackers had gotten into some of their data during the Fireworks AI incident.
Fireworks AI said its customer-facing platform had no more successful unauthorized access after Oct. 25. Uber said its access alerts failed to trigger because the attackers logged in with Fireworks’ trusted internal credentials. Uber gave affected individuals free credit monitoring and fraud assistance through Cyberscout, with 90 days from the notification letter to sign up.
Reuters said that, as of November 13, Uber had identified no misuse of the exposed information. The Fireworks incident was Uber’s second publicly known breach of 2026. In August, Uber investigated a separate Uber Freight incident after an extortion group claimed to have stolen data.

Join the discussion