Dragos Canary Test Exposes Three Breached AI Providers and Ignored Zero-Retention Pledges
Dragos seeded two million AI prompts with canary secrets over five months and found three breached providers, marketplace resellers logging prompts, and ZDR flags ignored as data reached a Chinese AI lab.
By Amara Osei
6 min read

What's News
- Dragos seeded ~2 million AI prompts with trackable secrets from July to November 2026 and found three hosting providers breached.
- The largest breach stole a couple hundred AWS keys in late October; stolen keys drove tens of thousands of AWS API calls.
- A Chinese AI lab scraped secret URLs from over half the prompts, including roughly two-thirds flagged Zero Data Retention.
- A related experiment by Lauren Proehl found median time from AWS key theft to first use was 37 minutes.
- Dragos recommends self-hosting sensitive models, with hardware costs from $6,950 for an NVIDIA DGX Spark to $570,000+ for a B300 server.
Security firm Dragos seeded nearly two million AI prompts with unique, trackable secrets over five months and caught three hosting providers being actively robbed, marketplace resellers reading customer prompts, and Zero Data Retention settings ignored across the board.
The experiment ran from July through November 2026. Dragos researchers submitted roughly two million prompts containing unique URLs, DNS names and AWS credentials to widely used AI services. Each secret existed in exactly one prompt, so any use of it proved someone had read that prompt. The team started with Thinkst Canarytokens, then built a dedicated service that generated a unique seed per prompt, derived hostnames and credentials from it, and ran the systems those credentials pointed to — DNS, web, git, SSH, databases and mail — with full logging. Every hit traced back to the exact prompt, model, router and privacy setting.
What did the canaries catch at the providers?
Dragos calls this the most serious finding: attackers stole data from three different hosting providers. The earliest incident dates to mid-July. None of the three providers knew they had been breached until Dragos told them.
The largest breach swept up a couple hundred AWS keys, each sent in exactly one request to one provider within short windows in late October. Every request Dragos sent during those windows was stolen. Some stolen keys came from requests the provider rejected before a model ever ran — the attacker was copying traffic at the provider's front door. In the other two breaches, some stolen requests carried Zero Data Retention (ZDR) flags.
Within hours of theft, attackers tested the credentials, enumerated S3 buckets, Secrets Manager and Parameter Store, and tried to enable hosted models in AWS Bedrock, create Bedrock API keys and add their own IAM users. The bait keys drove tens of thousands of AWS API calls.
One adversary appears to have hit two providers at once, using secrets from both almost immediately from a single rented server. The tooling evolved over three weeks: a script posing as Chrome on Windows, then Boto3 inside OrbStack, then the AWS CLI on Kali Linux. On November 1 the attacker added ARTEX, an LLM-driven penetration testing framework. Two user agents — artex-agent and artex-judge-r121 — do not appear in public indicator lists. Dragos does not name the providers, and the breaches have not been independently confirmed beyond Dragos's own telemetry.
How did marketplace resellers read the prompts?
No exploit was needed. Inference marketplaces let sellers list models at the lowest price and enable request logging, which exposes every prompt. Dragos caught multiple sellers doing exactly that.
In one case, two prompts sent through a single seller's offer were opened about 36 hours later from a home broadband connection, and someone began guessing usernames at the login page. Prompts for the same model through a different seller were never touched — the leak followed the seller, not the marketplace. In another, a Mac on a French home broadband line took passwords from two prompts, attempted hand-typed SSH root logins, then tested several AWS keys in a single session. One router was quietly forwarding some traffic to the marketplace, an undisclosed hop.
Did privacy settings actually protect anything?
No — and data reached a Chinese AI lab. A Chinese lab's systems fetched secret URLs from more than half the prompts Dragos sent to its models between July and October, including roughly two-thirds of requests marked "no data collection" or ZDR. The scraper typically appeared two to three days after the request. Several U.S. providers resell this lab's models, so prompts sent to U.S. providers triggered the same scraper — data flowed back to the lab despite customers choosing U.S. hosts to keep data in the U.S.
U.S. providers had their own problems. One kept crawling a secret URL more than two months after a single inference call. Another runs a scraper that returns days after each request regardless of router.
The strangest incident: 24 days after Dragos sent a prompt to a Chinese provider, the secret link was posted into a Feishu (Lark) chat channel and opened within two minutes from three desktops and three phones in Beijing and Shanghai. Dragos had disclosed only the first eight characters of the link to that provider, so whoever shared it was reading a stored copy of the full prompt more than three weeks after a ZDR-flagged request sent to a U.S.-based provider. Dragos states providers ignore settings "either intentionally or due to bugs in their processing flow" but hasn't distinguished between the two.
What did the route tracing reveal?
Dragos built a traceroute-like system for AI inference: an image URL in each request reveals every system that fetches it, and slightly broken requests expose nested error "envelopes" from each router. About 2,000 routes have been traced. Several hundred included unidentified systems; a handful sent images to companies not on the advertised path. One marketplace seller quietly added about 1,000 hidden tokens to every prompt; another fetched images through a shared consumer subscription account.
How did providers respond, and what's the corroborating evidence?
Dragos notified all affected providers. The first provider — the mid-July incident — initially thought it was a misconfiguration, then a partner data-processing issue, until investigation showed internal credentials were also compromised. As of late October, that provider had not publicly disclosed the incident, and stolen data continued to drive login attempts. One provider hit in late October published a preliminary security notice. For data-sharing issues, Dragos began notifying providers in early November and is working with about a dozen. Reactions are mixed: one router stopped forwarding to the marketplace and several providers opened investigations, but stolen keys and secret URLs were still being accessed months later. No provider has fully confirmed or disputed the findings beyond that preliminary notice.
A separate four-month experiment by Lauren Proehl, published November 19, 2026, seeded a fake developer box with canary AWS keys and SMTP credentials. Of 155 stolen AWS keys that were used, 72% went straight to SES/SNS messaging checks; median time from theft to first use was 37 minutes. Thieves (datacenter IPs) and users (residential broadband, 81% U.S.) were almost entirely different populations. LLMjacking — testing stolen keys for Bedrock access — grew from 12% of keys in August to roughly half in September and October, spread across about 20 unrelated networks. Credential checkers probed keys against fake AI endpoints designed to catch honeypots, and stolen material reached Telegram link previews within days.
Dragos recommends hosting models yourself for the most sensitive work — hardware costs run from $6,950 for an NVIDIA DGX Spark to $570,000-plus for an eight-GPU B300 server — going directly to a first-tier provider, verifying privacy settings actually change routing, avoiding reseller marketplaces, and seeding your own canaries, with a free Thinkst Canarytoken enough to start. Dragos notes the effort "felt a lot like what we do at Dragos with runZero; using small bits of data to provide a comprehensive view of networks from multiple perspectives" — a commercial interest in prompt-leak detection worth weighing when reading the results.
Source: epyhia.com
More from Amara Osei
Show full bio
Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.
2,422 articles