HD Moore designed an experiment that injected two million digital “canaries” – hidden tripwire secrets – into requests flowing through a sprawling network of AI services. The canaries, consisting of fresh AWS access keys, unused credentials, SSH passwords, unique URLs, DNS names and private‑git references, were routed via AI gateways such as OpenRouter and Vercel to inference‑engine providers including Anthropic and OpenAI. The test revealed that the ecosystem leaked highly sensitive data “like a sieve.”
Key Takeaways
- HD Moore planted two million previously unused digital canaries across AI routing services.
- The canaries contained AWS access keys, empty‑account credentials, SSH passwords, unique URLs, DNS names and private‑git references.
- AI routers and gateways (OpenRouter, Vercel) forwarded the canaries to inference‑engine providers such as Anthropic and OpenAI.
- The experiment demonstrated that the AI service ecosystem leaked the canaries at scale, described as “leaking like a sieve.”
| Metric | Value | Why it matters |
|---|---|---|
| Research lead | HD Moore | Identifies weaknesses in AI data handling practices. |
| Canaries planted | 2 million digital tripwires | Scale sufficient to stress‑test the entire AI supply chain. |
| Types of secrets | AWS keys, empty‑account credentials, SSH passwords, URLs, DNS names, private‑git refs | Represents a broad attack surface for potential exploitation. |
| Routing services used | OpenRouter, Vercel (AI routers/gateways) | Key choke points where user prompts are handed off to inference engines. |
Experiment Design and Methodology
The study injected the canaries directly into the request payloads that traverse AI routers. By embedding never‑used AWS access keys and other high‑value credentials, the experiment created a set of unique, traceable markers. Each canary was tied to an empty AWS account, ensuring that any observed usage could be unequivocally linked back to the AI service pipeline.
Findings: Data Leakage Across the AI Supply Chain
Analysis of the outbound traffic showed that the canaries surfaced in multiple downstream systems, confirming that the data was not confined to the intended inference engine. The leakage pattern was described as “like a sieve,” indicating that the AI ecosystem failed to contain or sanitize the embedded secrets. The breadth of the leak spanned providers such as Anthropic and OpenAI, demonstrating that the issue is not isolated to a single vendor.
The Role of AI Routers and Gateways
AI routers like OpenRouter and Vercel act as intermediaries, receiving massive volumes of user prompts and dispatching them to the most appropriate inference engine. Because these gateways handle the initial request parsing, they become a critical security boundary. The experiment shows that the current implementations allow embedded secrets to pass through unchecked, exposing both the routers and the downstream models to potential credential harvesting.
Implications for Inference‑Engine Providers
Providers such as Anthropic and OpenAI inherit the raw payloads from the routers. The observed leakage suggests that these engines either lack robust input sanitization or that their logging mechanisms inadvertently capture and expose the canaries. For organizations that rely on these services to process confidential data, the findings raise immediate concerns about compliance, data residency, and the risk of credential abuse.
Context: The Expanding AI Supply Chain
Modern AI applications increasingly depend on a layered architecture: end‑user interfaces send prompts to AI routers, which then forward them to a pool of inference engines. This modular approach accelerates development but also multiplies the attack surface. The experiment underscores that security controls must be enforced at every hop, not just at the final model inference stage.
Why it matters
The leak of two million fresh AWS keys and other high‑value credentials demonstrates a systemic weakness that could be exploited by malicious actors to gain unauthorized cloud access, exfiltrate data, or launch further attacks. Enterprises integrating AI services must reassess their data‑handling pipelines, enforce strict input validation, and consider end‑to‑end encryption to mitigate the risk of inadvertent secret exposure.
Frequently asked questions
Who conducted the canary‑planting experiment?
The experiment was designed and executed by researcher HD Moore.
What types of secrets were used as canaries?
The canaries included AWS access keys, functioning credentials tied to empty AWS accounts, SSH passwords, unique URLs, DNS names and references to private git repositories, all of which had never been used before.
Which AI routing services were involved?
OpenRouter and Vercel served as the AI routers/gateways that received and forwarded the canary‑laden requests.
Which inference‑engine providers received the leaked data?
Anthropic, OpenAI and other unnamed providers were among the inference engines that processed the requests.
What does the “leaking like a sieve” description imply?
It indicates that the AI ecosystem failed to retain or scrub the embedded secrets, allowing them to pass through multiple layers and become exposed.
Bottom Line
HD Moore’s deployment of two million digital canaries across AI routers and inference engines exposed a pervasive data‑leakage problem, with sensitive AWS credentials and other secrets surfacing throughout the AI supply chain. The findings highlight an urgent need for stronger sanitization and security controls at every stage of AI service orchestration.
The experiment serves as a warning that as AI integration deepens, the responsibility for protecting embedded secrets must be shared across routers, gateways and inference providers alike.
Reporting based on: original source.
Comments
Post a Comment