Security

AI Supply-Chain Prompt Leaks Expose Enterprise Security Controls Across Routers, Resellers, and Providers

Enterprises often treat prompt privacy as a vendor setting problem: enable Zero Data Retention, trust the provider, and assume sensitive input stays contained. The canary-token experiment reported by Ars Technica points to a weaker reality. Researcher HD Moore seeded nearly 2 million unique secrets into prompts sent through AI services and found evidence that sensitive data leaked across parts of the AI inference supply chain, including routers, resellers, and providers [Ars Technica].

Ars said the bait included AWS access keys, passwords, unique URLs, DNS names, and private git repository references, with each secret placed into exactly one prompt sent to a single provider [Ars Technica]. The operational lesson is not only that prompts can leak. It is that many teams may be drawing their trust boundary too narrowly, especially when traffic can pass through marketplaces and routing layers before reaching a model host, a risk category OWASP frames as an LLM supply-chain vulnerability [OWASP].

Why this is a supply-chain security story

Ars reported that prompts are encrypted in transit but decrypted at each hop inside the inference reseller network, where routers and providers can read and log them [Ars Technica]. Dragos, in Moore’s post on the experiment, similarly said AI routers can pass prompts through more than one router and then to the specific model host, and that every hop can see the full prompt and response [Dragos].

That makes the problem larger than a single provider breach. Ars reported findings that included evidence at at least three AI providers, and said multiple resellers in the inference marketplace logged prompts and somehow passed them along to people who later used the credentials [Ars Technica]. Dragos added that some marketplace sellers could siphon customer request data simply by offering the lowest price, enabling request logging, and waiting [Dragos]. That is the kind of hidden path dependence security teams should read alongside DailyOps coverage of safe AI browser testing environments, because both stories point to control weaknesses outside the polished user-facing product layer.

Prompt Path Security Watchboard

Every hop can become a disclosure point

Ars said Moore identified reseller leaks by sending requests to the same model through different sellers and observing which seller’s prompts led to credential reuse [Ars Technica]. Dragos said some stolen keys came from requests the provider rejected before a model ever ran, suggesting attackers were copying requests at the provider’s front door [Dragos].

ZDR is narrower than many teams assume

Ars reported multiple cases of AI providers not following privacy settings such as Zero Data Retention or “no data collection” [Ars Technica]. OpenRouter defines ZDR as provider-side processing without persisting prompts or responses afterward, and notes that tools, plugins, and other services have their own retention terms [OpenRouter].

Credential reuse is the operational alarm

Ars reported attackers used stolen bait AWS keys tens of thousands of times in AWS API calls, exploring S3 buckets, Secrets Manager, and Parameter Store, and trying to enable Bedrock and add IAM users [Ars Technica]. Ars also said some secrets sent to two hacked providers were used together almost immediately from the same rented server, suggesting the same adversary had compromised both [Ars Technica].

What broke the enterprise assumption

The core assumption that failed was that provider privacy labels automatically govern the whole route. Ars reported one Chinese AI lab fetched secret URLs from more than half the prompts sent to its models between July and October, including many marked “no data collection” or ZDR [Ars Technica]. Dragos likewise said prompts submitted to U.S.-based providers were in some cases transmitted to Chinese AI organizations [Dragos].

That does not mean every ZDR claim is false. It means ZDR is only one control, and even its strongest reading applies to provider-side retention rather than the whole enterprise request path [OpenRouter]. OpenRouter says guardrails can enforce ZDR-eligible endpoints, restrict providers, and apply prompt-injection defense and data loss prevention before a request is sent onward [OpenRouter]. In other words, enterprises need controls before the prompt leaves their side of the trust boundary, not just downstream policy promises.

What security teams should watch

  • Routing transparency: Inventory which routers, resellers, and model hosts can see prompt contents, not just the brand your developers think they are calling [Dragos] [OWASP].
  • Pre-send controls: Favor request-path controls such as DLP, provider restrictions, and explicit endpoint selection, rather than relying on retention language alone [OpenRouter].
  • Prompt canaries: Moore’s experiment shows why seeded secrets can reveal leaks that audit checklists miss, much like other AI privacy surprises now surfacing in the market, including the cross-conversation concerns reported around Meta’s Muse privacy alarms [TechCrunch].

DailyOps Insight

Operational takeaway: In AI systems, privacy settings are not the same as path controls. If prompts can traverse routers, marketplace resellers, rejected-request handlers, and model hosts, then the real security boundary is the full inference supply chain. Enterprises should treat prompt leakage as a supply-chain monitoring problem, not just a vendor policy review.

Knowledge Check

Your company enabled a provider’s Zero Data Retention option. What remaining risk from this story should still concern your security team most?
An attacker begins using credential bait from AI prompts to inspect S3 buckets and Secrets Manager. What is the clearest sign from this article that the leak became an active security incident?
If you want to reduce prompt leakage before traffic leaves your environment, which control direction best matches the article’s analysis?

Fast Takeaways

Sources