ZICQ
中 Log in / Sign up
Newsroom Industry & Trends #AI Routers #Prompt Data Leaks #Secret Management #Supply Chain Security #Canary Tokens

Two Million Digital Canaries Trace Secret Leaks Across AI Router Chains

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:HD Moore planted 2 million unique digital tripwires in prompts sent through a network of AI services. The markers included AWS access keys tied to empty accounts, SSH passwords, unique URLs and DNS names, and references to private Git repositories. By controlling the infrastructure that received leaked requests, Moore could trace how sensitive data moved through routers, gateways, and inference providers. The results show that a logical AI application can cross several administrative and techni


Two Million Digital Canaries Trace Secret Leaks Across AI Router Chains

HD Moore planted 2 million unique, previously unused digital markers in prompts sent through a network of AI services. By controlling the infrastructure that received any leaked requests, he could trace how sensitive information moved through routers, gateways, inference providers, and downstream services. The markers included AWS access keys, SSH passwords, unique URLs and DNS names, and references to private Git repositories.

The report cites OpenRouter and Vercel as examples of routers or gateways, and Anthropic and OpenAI as examples of large model providers. The available material does not establish that any named company mishandled data, so the findings should not be turned into provider-specific blame.

How the canary experiment localizes leaks

Each marker was assigned to one prompt and sent to one provider, while Moore controlled the endpoint infrastructure that would receive it. The resulting mapping links a secret marker to its prompt, origin, provider, routing decision, and receiving endpoint. When a controlled server observed a marker, investigators could use request identifiers and routing logs to reverse the exposure path.

The marker types served different detection purposes. AWS access keys tied to empty accounts retained identifiable account properties without placing valuable cloud assets at risk. SSH passwords could reveal that a prompt had been forwarded to an unexpected host. Unique URLs and DNS names exposed network destinations, while private Git repository references showed that code-access context had entered a downstream system.

High-entropy, one-time markers also reduce accidental collisions in natural-language traffic. Exact attribution becomes harder if a downstream service transcribes, truncates, encodes, or combines the marker with other secrets. Observing a marker also proves that data crossed its intended trust boundary; it does not, by itself, prove malicious exploitation by the recipient.

Why AI routers expand the attack surface

A conventional application often has a fixed backend and a relatively clear operational boundary. An AI router may instead select among inference engines according to price, task suitability, or availability. One logical request can therefore cross the application, gateway, router, model provider, tool connector, and multiple subprocessors.

The surfaces that require auditing include prompt and response logs, tracing systems, error payloads, caches, evaluation and observability platforms, support exports, tool arguments, and provider-side training or retention systems. The excerpt confirms supply-chain leakage but does not establish every one of these paths. A defensible audit must distinguish data that was logged, forwarded, or actually used by correlating marker receipts with request IDs and routing decisions.

Leakage probability is not simply the sum of independent per-hop risks. Shared logs, caches, and administrative permissions can make several risks highly correlated. Conversely, short-lived credentials, field-level redaction, and isolated sinks can cut multiple propagation paths at once. Controls therefore need to operate at verifiable boundaries rather than relying on broad contractual assurances that data will not be trained on.

Security benefits and engineering costs

  • Prompt redaction and structured tool calls: Detecting credentials and personal data before routing reduces model-side exposure, but regexes and semantic classifiers can remove useful context and add latency. High-confidence credential patterns should use allowlists and format validation rather than generic entity recognition alone.
  • Pinning trusted providers: Excluding sensitive requests from dynamic routing narrows the subprocessors involved, but gives up price optimization, capacity elasticity, and automatic model selection. A better compromise is tiered routing based on data classification.
  • Logging and observability: Full prompt logs improve debugging but replicate sensitive data. Safer audit records contain request IDs, providers, routing rationales, data classifications, policy versions, and field-level summaries. Raw content should be retained only for a short period under explicit approval.
  • Long-lived secrets versus short-lived capabilities: Putting AWS keys, SSH passwords, or Git credentials directly into prompts turns the model supply chain into a credential distribution system. A secret broker should issue narrowly scoped, expiring tokens only to approved tools after routing decisions are made.
  • No-retention commitments: Disabling training use, shortening log retention, and disclosing subprocessors are necessary, but contractual claims still need technical verification. Synthetic canaries can periodically probe whether the promised boundary is real.

Interpreting the two million markers correctly

Two million is a sample size, not a leakage rate. Calculating a rate requires a valid request denominator, observation window, provider and routing-policy combinations, retry and failover behavior, the location of secrets within prompts, and a definition of whether receipt alone counts as leakage or whether external use is required. Without those definitions, any percentage can be materially misleading.

Canary tokens are not identical to production secrets. They are often longer, more random, and stripped of business semantics, making them less exploitable after disclosure. Real credentials may be obscured by natural-language context and then reused through encoding, scripts, or manual operations. A credible evaluation should report raw marker occurrence, recovery after transformation, effective-credential use, and independent attribution at every stage.

Reproduction also requires negative controls without secret injection, repeated trials, routing-configuration snapshots, per-stage latency and failure rates, confidence intervals, and complete ablations. Without them, a large sample proves that the experiment happened; it does not establish whether the router, logging layer, or model provider is the dominant source of risk.

Production recommendations

  1. Map the complete data flow across applications, routers, gateways, models, and tool connectors, recording the data class, retention period, region, and responsible entity at every hop.
  2. Run DLP checks before any prompt enters a router. Remove long-lived credentials from context and replace them with runtime-issued, short-lived, least-privilege capabilities.
  3. Apply provider allowlists for sensitive requests, require no-training and retention controls, and prohibit unreviewed cross-region or cross-subprocessor routing.
  4. Make audit logs metadata-first, using immutable fingerprints instead of raw secrets by default. Gate any temporary plaintext access for debugging.
  5. Continuously validate end-to-end isolation with unique synthetic markers, and calculate occurrence rates separately by provider, routing policy, and time window rather than reporting only one aggregate number.
  6. Maintain rapid credential revocation and rotation. If a marker appears at an unauthorized endpoint, responders should be able to identify the request, locate the responsible node, and initiate the provider's incident process.

The experiment's main value is not a claim that every AI service will immediately abuse secrets. It turns the assumption that prompts leave the application process into a measurable supply-chain risk. An AI router is not a security boundary by itself; without stronger secret handling, data governance, and observability, the cost and flexibility gains of dynamic routing may come at the price of a much larger sensitive-data exposure surface.


Source: Ars Technica AI (2026-11-20)

— END —

Tags: #AI Routers #Prompt Data Leaks #Secret Management #Supply Chain Security #Canary Tokens

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…