Security researchers at Dragos planted 2 million canary tokens in AI inference requests, uncovering a supply chain where credentials and prompts are routinely harvested by third-party intermediaries. The researchers ran the tests from July through November 2026 and identified at least three compromised providers where attackers were actively stealing user data. None of the three knew they had been compromised until Dragos notified them, according to reports published by The Decoder and Ars Technica.
The canary test and what it found
To test for potential leakage, Moore assembled a collection of Amazon Web Services access keys tied to empty accounts, SSH passwords, and unique URLs, DNS names, and private git repository references. None of the secrets had ever been used before, and he controlled the infrastructure that received any requests using the credentials. He seeded each secret into exactly one prompt sent to a single provider. In the event any of the secrets were leaked, these “canaries” would allow the team to identify each link in the chain responsible for the breach. For developers building AI applications, the results mean that routing prompts through third-party brokers exposes sensitive data to any provider in the pool, regardless of the privacy guarantees advertised by the primary platform.
The results showed a system leaking sensitive information at scale. In the largest incident, attackers harvested hundreds of AWS keys, including some from requests the provider had rejected before processing. The attackers tested the stolen keys within hours, searched S3 storage, and attempted to establish persistent access. One attacker even used ARTEX, an AI-powered attack tool.
Marketplace resellers did not need to exploit software vulnerabilities to access this data. Simply listing a model at the lowest price and turning on logging was enough to capture customer requests. In one case, an individual on a residential internet connection in France used SSH passwords pulled from intercepted requests to attempt logins.
The intermediary routing crisis
The exposure highlights the risks of AI routers and gateways, which act as brokers between end users and inference engines. Instead of calling Anthropic or OpenAI directly, developers hire routers to find the most suitable model based on criteria like price or task fit. Requests are routed through a pool of providers, sometimes across several hops, and each hop can read the entire request and response.
In this case, providers including Novita, AtlasCloud, and StreamLake sat between OpenRouter and model developer MiniMax. Together runs its own infrastructure, and Phala has its roots in Web3. Following the research, OpenRouter responded by removing three providers from its platform and revoking Zero Data Retention approval for 125 endpoints. The company also emailed affected users, warning that third parties may have read requests sent to MiniMax models. The exposure involved prompts routed through resellers AtlasCloud, Novita, and StreamLake.
Illustrative image: geralt / Pixabay
Affected providers have published preliminary security reports on the incident. Dragos warns companies against sending requests containing credentials to intermediaries, recommending that organizations control how their AI requests are routed 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," HD Moore writes.
A recent analysis by the Oxford China Policy Lab showed how Chinese proxy services route requests to Western AI models through servers abroad. Those services can also read the requests in transit. It remains unclear whether the providers Dragos examined share any personnel with those proxy networks. The methods, however, follow the same basic pattern: intermediaries sit between users and model developers, view requests, and pass data to third parties without user knowledge. Data ends up with foreign actors even when users never intended to send it there, and privacy promises break down along the supply chain.
Privacy settings and data retention failures
The research also exposed widespread failures in data retention policies. Privacy settings like Zero Data Retention were frequently ignored. Dragos reported that a Chinese AI lab accessed secret test URLs in more than half of all requests, even when those requests went through US providers.
Even major US providers broke their own policies. One provider was still crawling a test URL more than two months after a single request. In another case, a secret test link appeared in an internal chat on Feishu (known outside China as Lark) 24 days after the request. The link was opened within two minutes on three computers and three phones in Beijing and Shanghai.
The findings suggest that developers relying on third-party routing to cut costs are effectively outsourcing data access to any provider in the pool, regardless of the privacy guarantees advertised by the primary platform.
Sources
- the-decoder.com: Security researchers show how cheap AI inference providers steal keys and snoop on users
- arstechnica.com: Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.
- techcrunch.com: Meta’s AI agent Muse is mixing up conversations and could be exposing private data, report says | TechCrunch
PNEUMETRON EDITORIAL
Pneumetron's pipeline pairs AI-assisted drafting with human editorial review before publishing — our goal is to make staying informed easier, not to replace real reporting.




