Actively exploited Security news

Attackers read prompts inside three AI inference providers for months; OpenRouter drops them

Over five months, Dragos researchers sent roughly two million prompts laced with unique canaries to AI services and watched where they surfaced. Attackers were stealing requests from inside three inference providers, including requests flagged Zero Data Retention. On 22 November 2026 OpenRouter off-boarded three providers and pulled ZDR status from 125 endpoints.

Kyberkit editorial team · · event date · reading time 3 minutes · For admins and analysts

Actively exploited: AI inference providers

Key points

  • Dragos (HD Moore, 19 Nov 2026) planted unique URLs and decoy credentials in prompts. By tracing which ones were used, it uncovered theft at three providers, the earliest dating to mid-July.
  • Zero Data Retention did not help: an attacker inside the provider's infrastructure reads requests before any retention decision is made.
  • OpenRouter says its own systems were not breached. It removed three unnamed providers, pulled ZDR from 125 endpoints and notified affected customers.
  • Phala has publicly disclosed an intrusion in which an attacker could read DeepSeek V4 Flash 0731 request contents between 21 Oct and 8 Nov.

What happened

Between July and November 2026, a Dragos team led by HD Moore sent roughly two million prompts to widely used AI services. Each carried unique URLs and decoy credentials: AWS keys tied to empty, fully logged accounts, plus Git, SSH and database logins. Whenever someone tried a decoy, it could be traced back to the provider, model and router the prompt went through.

Dragos published the results on 19 November: attackers were stealing data from inside three inference providers. In the largest breach, every request sent during several windows in late October was taken, including requests the provider rejected before a model even ran. In the other two, stolen requests included ones flagged Zero Data Retention.

What the attackers did with it

Stolen AWS keys were tried within hours: listing S3 buckets, searching Secrets Manager and Parameter Store, trying to enable Bedrock models and adding IAM users. The decoys drew tens of thousands of AWS API calls. Dragos says one actor controlled two providers at once and, from 1 November, brought in the offensive AI framework ARTEX.

Dragos also describes quieter leaks: marketplace sellers who switched on request logging and could read prompts, a router that forwarded some traffic to a marketplace without disclosing it, and a Chinese AI lab whose systems fetched secret URLs from more than half the prompts sent to its models, including about two-thirds of those marked "no data collection".

How providers are responding

  • OpenRouter said on 22 November that its own systems were not breached. It off-boarded three providers, removed ZDR eligibility from 125 endpoints and notified affected customers. It has not named them.
  • Phala's preliminary report says an attacker chained a legacy internet-facing service to its server management interfaces. Request contents for DeepSeek V4 Flash 0731 may have been readable from 21 October to 8 November.

How to check for compromise

Dragos lists two ARTEX user agents it did not find in public indicator feeds. Search CloudTrail and access logs for:

artex-agent
artex-judge-r121

If AWS keys ever went into a prompt, review CloudTrail for unexpected ListBuckets, GetSecretValue and GetParameter calls, Bedrock model enablement and new IAM users. Our IOC extractor pulls addresses out of log excerpts.

What to do

  1. Rotate secrets that may have passed through a router, marketplace or smaller provider in a prompt, especially if OpenRouter or Phala has notified you.
  2. Check ZDR status on your endpoints and pin providers yourself instead of relying on automatic routing.
  3. Avoid cheap marketplace listings from unknown sellers.
  4. Turn on secret filtering where your provider offers it, and seed your own decoy credentials in agent context; treat any use as an incident.

Timeline

  • mid-July 2026 – earliest theft Dragos traced
  • 21 Oct – 8 Nov – attacker inside Phala
  • late October – largest breach (a couple of hundred canary AWS keys)
  • 19 Nov – Dragos publishes its research
  • 22 Nov – OpenRouter removes three providers

FAQ

I use ChatGPT or Claude directly from the vendor. Does this affect me?

The findings mainly concern smaller hosting providers, marketplaces and routers that requests pass through. Dragos notes that going direct to a first-tier provider reduces, but does not eliminate, the risk. The safest option is to keep passwords and keys out of prompts altogether.

What is Zero Data Retention?

A setting under which the provider promises not to store the prompt or response. It is a contractual promise, not a technical control, and cannot stop someone who controls the serving machine from reading the data.

How do I know whether my endpoint lost ZDR status?

Check your OpenRouter settings for the providers and endpoints you allow and whether they are still marked ZDR. OpenRouter removed 125 endpoints from ZDR and says it will restore eligibility only after validation.

Sources

  1. Dragos - Three breaches, endless leaks: Your AI supply chain needs canaries (19 Nov 2026)
  2. OpenRouter - Data processing incidents on inference providers (22 Nov 2026)
  3. Phala - November preliminary incident report (10 Nov 2026)

Spotted a mistake? Let us know and we will correct it openly.