Dragos says AI inference providers and resellers leaked keys and prompts
Using about 2 million trackable prompts, Dragos says it found three breached hosting providers, resellers reading requests and ignored privacy settings.
Key takeaways
- Dragos sent about 2 million AI prompts with trackable fake secrets to AI services from July through November 2026.
- It says attackers were stealing data at three hosting providers, none of which knew until Dragos told them.
- Some marketplace resellers could read customer prompts just by enabling request logging, Dragos says.
- From July to October, a Chinese AI lab opened secret URLs in over half of the prompts sent to its models, including many marked Zero Data Retention, according to Dragos.
- OpenRouter removed three providers and revoked Zero Data Retention approval for 125 endpoints, The Decoder reports.
Security firm Dragos says it found attackers stealing data from three AI hosting providers, marketplace resellers reading customers' prompts, and providers ignoring privacy settings, after sending about 2 million test prompts containing trackable fake secrets to AI services. The Decoder also reported the findings.
From July through November 2026, Dragos researchers seeded prompts with unique URLs and canary credentials that no legitimate system would use, then watched for anyone touching them, the company wrote. Any use showed that someone along the way had seen the prompt.
What did Dragos find?
Breached providers. Attackers were stealing data from three hosting providers, and none knew until Dragos told them, according to the company. Dragos says real customer requests processed alongside its bait were almost certainly copied too. In the largest case, a couple hundred of Dragos' AWS keys were taken, including from requests the provider had rejected before a model ran, which Dragos says shows the copying happened at the provider's front door. The attackers tested keys, sometimes within hours, searched storage and tried to keep access; in total the planted keys drew tens of thousands of calls to AWS APIs. One attacker later added ARTEX, an AI-driven penetration testing tool, though Dragos says it did very little in this case beyond a handful of calls to check that keys were still live.
Resellers reading prompts. Some sellers on an AI inference marketplace could read and keep customer prompts simply by offering a model at the cheapest price and switching request logging on, without exploiting any flaw, Dragos says. It adds that the leak followed individual sellers, not the marketplace as a whole. In one case, someone on a home connection in France used passwords taken from prompts to try to log in to Dragos' test servers over SSH.
Privacy settings ignored. Systems belonging to a Chinese AI lab opened the secret URLs in over half of the prompts Dragos sent to that lab's models between July and October, including about two-thirds of requests marked "no data collection" or Zero Data Retention, Dragos says, and prompts sent to U.S. providers reselling its models triggered the same scraper. A large U.S. provider kept crawling one secret URL for over two months after it appeared in a single request to the provider's search endpoint, Dragos says. Separately, a link from a prompt appeared in a Feishu (Lark) chat 24 days later and was opened on six devices in Beijing and Shanghai, Dragos says. Dragos did not name the lab or the providers.
How did the industry respond?
OpenRouter, which routes requests to models from many providers, removed three providers from its platform and revoked Zero Data Retention approval for 125 endpoints, partly because of the findings and partly as a precaution, The Decoder reports. It also warned affected users that third parties may have read requests sent to MiniMax models; those requests passed through intermediary providers that sat between OpenRouter and MiniMax, The Decoder reports. Neither Dragos nor The Decoder identifies the Chinese lab. Some affected providers, among them Together, Phala and Novita, have put out early security reports, but others have not yet, The Decoder reports.
Dragos says reactions have been mixed. The router that had been quietly passing traffic to the marketplace told Dragos it had stopped, and several providers have started their own investigations. The provider breached first, in mid-July, initially thought the problem was a misconfiguration of internal tooling and has not publicly disclosed the incident, according to Dragos, which says it is working with about a dozen providers.
What should companies do?
Dragos advises against sending credentials in prompts that pass through intermediaries, and recommends that organizations decide for themselves how AI requests are routed, review privacy settings and plant their own canary tokens. "None of this means that you should stop using AI services. It does mean that where your prompts go is a decision, and if you don't make it, the lowest price will make it for you," wrote HD Moore, Dragos' chief research officer, in the company's post.
Questions readers are asking
How did Dragos detect the leaks?
It put unique URLs and fake credentials into about 2 million prompts and watched for anyone using them. Any use showed that someone had seen the prompt.
Which providers were affected?
Dragos did not name them. The Decoder reports that OpenRouter removed three providers and that Together, Phala and Novita have put out early security reports.
How can companies protect their prompts?
Dragos recommends not sending credentials through intermediaries, controlling how AI requests are routed, checking privacy settings and planting canary tokens.
Sources
- #dragos
- #ai-supply-chain
- #data-leak
- #openrouter
- #canary-tokens
- #ai-inference
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
Cybersecurity2 min
BADBOX 3.0 botnet hit 40 billion fraudulent ad requests a day, HUMAN says
Researchers say cheap Android-based TV boxes made in China came pre-infected, played hidden video ads and rented out owners' internet connections. HUMAN and Google disrupted it.
Cybersecurity2 min
Microsoft says Entra Connect Health flaw exposed admin emails
A bug in the identity monitoring service's API, present since July, gave outsiders read-only access to tenant metadata. Microsoft says passwords and tokens were not exposed.
Cybersecurity2 min
FBI reportedly investigating breach of its NCIC crime database
The national database police use for warrants and background checks was reportedly breached, the second reported FBI intrusion in recent months. The USDA was also hit this week.
Cybersecurity2 min
Canada warns servers running AI hacking tool ARTEX target its organizations
The Canadian Centre for Cyber Security says AI-enabled offensive security frameworks are spreading, and that servers running ARTEX have targeted Canadian organizations for months.