AI routing platform OpenRouter has removed three third-party service providers after cybersecurity researchers uncovered unauthorised access to artificial intelligence prompts and sensitive credentials across multiple AI hosting services. The investigation revealed that some requests marked for Zero Data Retention, a privacy setting intended to prevent prompt storage, had been accessed by attackers or processed beyond their expected destinations.
The findings emerged from a months-long investigation by industrial cybersecurity firm Dragos, which tested approximately two million AI prompts between July and November 2026 using specially planted credentials and tracking links.
Following the disclosures, OpenRouter removed three providers from its platform, revoked Zero Data Retention eligibility for 125 AI endpoints and notified potentially affected customers.
The company maintained that its own infrastructure had not been breached and that the incidents involved third-party providers handling customer requests.
FCRF Launches Certified Chief Risk Officer (CRO) Program for India’s Risk Leaders
How Did Researchers Discover Months of AI Prompt Theft?
The investigation relied on a technique involving canary tokens, which are specially created credentials and web links designed to reveal when supposedly private information has been accessed.
Researchers inserted unique credentials and tracking links into approximately two million AI requests sent through different inference providers.
Each planted credential appeared in a specific request, allowing investigators to identify where the information had been exposed when someone subsequently attempted to use it.
The research uncovered three compromised AI hosting providers, with the earliest suspected intrusion dating back to mid-July.
In one incident, researchers found that hundreds of planted Amazon Web Services (AWS) credentials had been accessed after being submitted through an AI provider.
The attacker tested some credentials within hours of obtaining them and attempted to access cloud storage, AWS Secrets Manager and other services.
Investigators also observed attempts to create additional cloud credentials and establish persistent access through AWS Identity and Access Management accounts.
The activity generated tens of thousands of AWS API calls.
Researchers determined that certain credentials had been stolen before the corresponding requests reached the AI models, suggesting that attackers were intercepting traffic at an early stage of the processing infrastructure.
Evidence also indicated that one attacker may have compromised two separate providers and used the same rented server to test credentials obtained from both.
On November 1, the attacker used an AI-enabled penetration testing framework called ARTEX to check whether previously obtained credentials remained active.
The findings raised concerns that genuine customer requests processed alongside the research prompts could also have been exposed.
Why Did Zero Data Retention Fail to Protect Some AI Requests?
A significant finding concerned the handling of prompts submitted under Zero Data Retention policies.
These settings are intended to prevent AI providers from storing user prompts, particularly when organisations submit sensitive business information through AI services.
However, the researchers found instances in which prompts marked for restricted retention were accessed, forwarded to additional providers or processed in ways that contradicted customers’ expectations.
In one case, sellers operating through an AI inference marketplace offered low-cost model access while enabling logging of customer requests.
Researchers observed that individuals operating some of these services accessed planted credentials and attempted to use them.
One operator, traced to a residential internet connection in France, reportedly attempted to access systems using passwords extracted from test prompts.
The investigation also identified instances in which AI requests were routed through additional services without the customers being informed.
Tracing approximately 2,000 routing paths revealed several hundred unidentified systems involved in processing requests.
Researchers also examined activity associated with an unnamed Chinese AI laboratory. They found that secret tracking links embedded in more than half of the test prompts sent to its models between July and October had been accessed.
The activity included requests marked for restricted data collection or Zero Data Retention.
In another incident, a tracking link from a prompt appeared in a workplace messaging conversation 24 days after the original request. The link was subsequently accessed from multiple devices in Beijing and Shanghai.
The findings indicated that sensitive AI prompts could remain accessible through downstream systems even when users believed that their information would not be retained.
What Have OpenRouter and the Affected AI Providers Disclosed?
OpenRouter said it had removed three providers and withdrawn Zero Data Retention eligibility from 125 endpoints following the investigation.
The company explained that the affected endpoints included services identified by researchers and others removed as a precaution.
Providers seeking to regain their previous privacy status will be required to provide additional assurances and undergo validation.
OpenRouter is also developing continuous testing methods based on the researchers’ approach and reviewing network routes to identify situations where requests may be forwarded to undisclosed third parties.
Several AI infrastructure providers have separately disclosed security incidents involving customer prompts.
Phala confirmed that an attacker had gained control of an inference server and could access request information while it was being processed.
The company said the incident was contained on November 8 and advised customers whose requests to the DeepSeek V4 Flash 0731 model passed through its service between October 21 and November 8 to treat sensitive information in those requests as potentially exposed.
Customers were advised to rotate passwords, API keys, tokens and other credentials submitted during the affected period.
Together AI disclosed that an unauthorised party had downloaded stored samples of customer prompts after an AI coding agent accidentally committed a credential to the company’s public GitHub repository.
The affected samples involved requests to the MiniMax M3, GLM-5.2 and Kimi K2.6 models during June and July 2026.
According to the company, the attacker searched the downloaded material for credentials associated with external services and attempted to use them.
Together AI said it had no reason to believe the information had been distributed further and understood that the downloaded data had been deleted. The company has since made Zero Data Retention the default setting for new accounts.
Another provider, Novita, acknowledged that a routing configuration issue had caused some requests intended for MiniMax models to be forwarded to MiniMax’s official API beyond the processing arrangements expected by customers.
The incidents have raised concerns about the security of increasingly complex AI supply chains, where a single user request may pass through multiple hosting providers, routing services and infrastructure operators.
Researchers warned that privacy settings alone cannot prevent exposure when an attacker has already compromised the infrastructure processing requests.
They recommended that organisations carefully select their AI providers, verify how requests are routed and avoid submitting sensitive credentials or confidential information unless the security and retention arrangements are understood.
The investigation highlights the risks businesses face when relying on third-party AI services to process confidential information, particularly where data handling practices are difficult to verify and security failures at one provider can expose requests originating from multiple customers.
About the author — Suvedita Nath is a science student with a growing interest in cybercrime and digital safety. She writes on online activity, cyber threats, and technology-driven risks. Her work focuses on clarity, accuracy, and public awareness.
