As reported by The Hacker News, Dragos researchers have disclosed that three AI inference providers silently lost customer prompts for months — including requests explicitly flagged for Zero Data Retention. OpenRouter has since removed all three providers and disabled 125 endpoints from eligibility for that privacy setting. The findings expose a structural trust problem in the AI inference economy that no privacy flag can fix.
The Trust Boundary Is a Fiction
The most uncomfortable takeaway is not that three providers were compromised. It is that every customer who selected Zero Data Retention believed a configuration toggle could enforce a contractual promise across infrastructure they cannot inspect. HD Moore's observation —
a privacy flag doesn't help when the attacker is already inside— is the entire story in one line. Zero Data Retention is a policy, not a control. The moment an inference provider's ingress, load balancer, or logging layer is compromised, the prompt is already copied before the policy is ever evaluated.
Dragos seeded roughly two million canary prompts with unique URLs and credentials between July and November. The technique is elegant: a planted secret can only be discovered by reading the prompt that carries it, so any use of that secret proves the prompt was read. This is the same logic that makes canary tokens effective in credential monitoring and honeytoken deployments — and it maps cleanly onto AI traffic where traditional DLP never sees the payload.
What Was Actually Stolen
The largest breach involved several hundred AWS keys planted in prompts sent during short windows in late October. Every request in those windows was exfiltrated, including some the provider rejected before any model ran — indicating the attacker sat at the ingress layer, not deep in the inference stack. Within hours, those keys were used to enumerate S3 buckets, query Secrets Manager and Parameter Store, and attempt creation of Bedrock API keys and persistent IAM users. One operator pivoted across two providers from a single rented server and on November 1 introduced ARTEX, an AI-driven penetration testing framework, solely to validate that stolen keys remained live.
The second vector is equally important and easier to overlook: marketplace sellers listing models at the lowest prices, enabling request logging, and harvesting whatever flowed through. One operator on a home Mac in France exfiltrated credentials from just two prompts. This is not a sophisticated breach — it is a business model built on the assumption that customers will not audit their providers.
Why This Matters Beyond OpenRouter
OpenRouter is a convenient focal point, but the risk is systemic. Any organization routing LLM traffic through a gateway, broker, or third-party inference provider inherits the provider's security posture as its own. The supply chain for AI inference is opaque, rapidly commoditized, and price-driven — exactly the conditions that attract low-cost malicious operators and make persistent compromise easy to disguise as cheap hosting.
Shield53 Recommendations
Immediate Actions
- Audit every AI inference vendor. Require evidence of logging controls, ingress monitoring, and incident detection. If the provider cannot demonstrate detection capability, treat all prompts sent as potentially exposed.
- Deploy canary prompts. Seed routine AI traffic with unique, monitored credentials and URLs. Any activation indicates prompt theft. This is low-cost, high-signal, and currently the best available control.
- Never put live secrets in LLM prompts. Treat the prompt field as untrusted output. Use scoped, ephemeral, or synthetic credentials for any workflow that requires model access to secrets.
- Block the ARTEX indicators. Add the
artex-agentandartex-judge-r121user agents to detection rules. Dragos notes these do not appear in public indicator lists. - Prefer self-hosted or enterprise-tier inference. For sensitive workloads, the cost delta between a marketplace provider and a vetted enterprise tier is less than the cost of one leaked AWS root key.
Strategic Actions
- Build AI vendor risk into procurement. Security questionnaires for AI providers must cover ingress logging, breach detection, and data retention enforcement — not just SOC 2 checkboxes.
- Monitor cloud credentials sent to AI services. Treat any AWS key that has ever appeared in an LLM prompt as compromised. Rotate immediately and review CloudTrail for enumeration patterns.
- Pressure gateways for provider transparency. OpenRouter responded well here, but customers should demand provider-level breach notification as a contractual term, not a courtesy.
The AI inference supply chain is behaving like every other software supply chain before it learned its lessons the hard way. The difference is that the payload in transit is natural language — which means secrets, internal context, proprietary code, and business logic all travel together in a format no traditional DLP tool parses well. Organizations that treat prompt traffic as untrusted data flowing to an untrusted endpoint are the ones who will sleep through the next breach. Everyone else is already in it.