Dragos Research Finds Three AI Provider Breaches, Prompt Theft, and Ignored Zero Data Retention Settings in Canary Token Test Dragos published research in November 2026 reporting that a five-month experiment seeding roughly two million AI prompts with unique canary tokens uncovered three AI provider breaches, unauthorized prompt access by marketplace resellers, and failures to honor Zero Data Retention settings. The largest breach involved "a couple hundred" of the seeded AWS keys, all stolen within "a few short windows in late October," after which attackers made tens of thousands of AWS API calls, enumerated S3 buckets, and attempted to enable AWS Bedrock models and create IAM users. Dragos Chief Research Officer HD Moore led the work, which also traced roughly 2,000 inference routes and found several hundred including unidentified systems; the providers are not named and no independent confirmation of the breaches has been reported. Dragos Research Finds Three AI Provider Breaches, Prompt Theft, and Ignored Zero Data Retention Settings in Canary Token Test Security firm Dragos published research in November 2026 claiming that a five-month experiment seeding nearly two million AI prompts with unique canary tokens uncovered three provider breaches, unauthorized prompt access by marketplace resellers, and widespread failures to honor Zero Data Retention ZDR and other privacy settings. The providers involved are not named. No independent confirmation of the breaches has been reported in public sources. One of the late-October providers has published a preliminary security notice, Dragos says; the mid-July provider has not publicly disclosed the incident, and the status of the third provider is unknown. The experiment From July through November 2026, Dragos researchers submitted approximately two million prompts containing unique, trackable secrets — URLs, DNS names, and AWS credentials — to widely used AI services. The tokens were generated so that any use would indicate the prompt had been seen by someone in the inference chain. The team began with Thinkst Canarytokens and later built a dedicated service that derived unique hostnames, URLs, and credentials from a per-prompt seed, running the services those credentials pointed to DNS, web, git, SSH, databases, mail and logging every interaction. AWS keys were real, tied to empty accounts with full CloudTrail logging. To map the inference supply chain, Dragos also built what it describes as a "traceroute for AI inference": embedding an image URL that reports back the fetcher's IP and client, and sending slightly broken requests whose nested error envelopes reveal each routing hop. The firm says it traced roughly 2,000 routes, finding several hundred that included unidentified systems and a handful that sent data to companies not on the advertised path. Dragos Chief Research Officer HD Moore, creator of Metasploit, leads the team. Dragos sells security products and notes this work resembles its runZero network visibility offering. Finding 1: Three provider compromises Dragos reports observing activity indicating that data from at least three AI hosting providers was exposed. The earliest incident dates to mid-July. Dragos says none of the three providers knew of the breach until notified. The largest breach involved "a couple hundred" of the seeded AWS keys. Each key was sent in exactly one request to one provider, all within "a few short windows in late October," and every request sent during those windows was stolen. Some stolen keys came from requests the provider rejected before a model ran, which Dragos interprets as the attacker copying requests at the provider's front door. In the other two breaches, some stolen requests carried ZDR flags. "A privacy flag doesn't help when the attacker is already inside," the post states. Once attackers had the credentials, they tested them within hours, enumerated S3 buckets, searched Secrets Manager and Parameter Store, and attempted to enable AWS Bedrock models, create Bedrock API keys, and add IAM users. The bait keys were used for tens of thousands of AWS API calls. Dragos says one adversary appears to have compromised two providers simultaneously: secrets sent to both were used together, almost immediately, from a single rented server. The attacker's tooling shifted over three weeks from a script mimicking Chrome on Windows, to Boto3 in OrbStack, to the AWS CLI on Kali Linux. On November 1, the attacker added ARTEX, an LLM-driven penetration testing framework. In Dragos's environment, ARTEX made only a handful of calls to confirm already-held keys were live. Two user agents it sent — artex-agent and artex-judge-r121 — do not appear in public indicator lists Dragos could find. The provider breached in mid-July initially thought the activity was a misconfiguration of internal tooling, then a partner data-processing issue, before finding evidence that an attacker had compromised its internal credentials tool. As of late October, that provider had not publicly disclosed the incident, and stolen data continued to generate login attempts. For the late-October intrusions, Dragos says it reached out once third-party access was confirmed, and one of those providers has published a preliminary security notice. Finding 2: Marketplace resellers reading prompts without exploits The second finding required no vulnerability. Dragos describes AI inference marketplaces where sellers list models at the lowest price, enable standard request logging, and collect whatever passes through. The firm says it caught multiple resellers viewing and retaining customer prompts this way. In one case, two prompts sent through a single seller's offer were opened roughly 36 hours later from a home broadband connection; someone manually guessed usernames at the login page. Prompts for the same model routed through a different seller were never touched, indicating the leak followed the seller, not the marketplace as a whole. In another, a Mac on a French home broadband line extracted passwords from two prompts and attempted hand-typed SSH root logins, then tested several AWS keys in a single session. One router was found quietly forwarding traffic to the marketplace — an undisclosed hop. Dragos explains that routers often chain: a request may pass through Vercel, then OpenRouter, then another provider, then the model host. Every hop sees the full prompt and response. Some sellers obtain cheap capacity through consumer-plan subscriptions, startup credits, or, in "shadier marketplaces," stolen AWS credentials or consumer API keys. Finding 3: Privacy settings and retention policies ignored The third finding involved no stolen credentials. Secret URLs in prompts were scraped repeatedly, and referenced git repositories were targeted for cloning, by companies that were never supposed to see the prompts — often in direct conflict with the request's privacy settings. The clearest example involves a Chinese AI lab. Between July and October, its systems fetched the secret URLs from more than half of the prompts sent to its models, 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. hosting providers resell this lab's models; prompts sent to those U.S. providers triggered the same scraper, meaning data flowed back to the lab despite customers choosing U.S. providers to keep data in the U.S. Dragos also describes U.S. providers with their own retention issues: one continued crawling a secret URL more than two months after a single inference call to its search endpoint; another runs a scraper that returns days after each request regardless of router. The most unusual incident: 24 days after a prompt was sent to a Chinese provider, the secret link appeared in a Feishu Lark chat channel and was opened within two minutes from three desktops and three phones in Beijing and Shanghai. Dragos had disclosed only the first eight characters of that link to the provider, so whoever shared it was reading a stored copy of the full prompt more than three weeks after the request — which Dragos says was sent to a U.S.-based provider with the ZDR flag. The explainx.ai summary notes that Dragos's own write-up is not fully consistent on which provider received that request \ 2\ https://explainx.ai/blog/dragos-hd-moore-two-million-canary-prompts-three-provider-breaches-zdr-ignored-ai-routers-2026 . What is verified and what is not From Dragos's post: the experiment design and duration; the three categories of findings with specific examples; that it is working with about a dozen providers; that one provider published a preliminary security notice; that one router told Dragos it stopped forwarding to the marketplace. Not independently confirmed: the breaches themselves, since providers are not named; the share of prompts affected beyond Dragos's own samples; whether the ignored ZDR settings were intentional or bugs Dragos says "either intentionally or due to bugs" ; the identity of the Chinese AI lab or the U.S. providers involved; the Feishu chat incident's chain of custody. No provider statements disputing or confirming the findings have been located in public sources. Dragos acknowledges that providers "often ignore these settings, either intentionally or due to bugs in their processing flow." The firm also sells security products and frames this research as akin to its runZero work, a commercial interest in prompt-leak detection. Provider responses and ongoing exposure Reactions have been mixed. The router forwarding traffic to the marketplace said it stopped. Several providers have responded and started investigations. However, Dragos reports that stolen keys and secret URLs continue to be accessed months after the initial incidents. For less severe issues like data sharing without credential abuse, Dragos began notifying providers in early November, providing lists of stolen requests and unique strings to search their logs. Mitigations Dragos recommends 1. Self-host for the most sensitive work. Hardware costs: Mac Studio 512 GB at $22,299, NVIDIA DGX Spark at $6,950, DGX Station at ~$105,000, eight-GPU B300 server at $570,000+. 2. Go directly to a first-tier provider. One company, one contract, one privacy policy — though breaches and data-sharing agreements can still leak data. 3. Use routers with care. Choose upstream providers yourself, verify privacy settings actually change routing, avoid reseller marketplaces some capacity comes from stolen keys; some accept only cryptocurrency . 4. Seed your own canaries. Add a fake credential to agent and pipeline contexts; treat any use as an incident. A free Thinkst Canarytoken is a starting point. 5. Enable secret filtering. Some providers scrub known credential formats; it helps but won't catch sensitive data not shaped like an API key. What remains unknown The providers' identities, their formal responses, whether law enforcement or regulators have been engaged, and whether the observed scraping and chat-sharing reflect policy, bugs, or individual employee actions. Dragos says it would not be surprised to see the number of providers it is working with expand. Until providers are named and respond publicly, the findings remain Dragos's observations from a controlled experiment — valuable signal, but not yet independently verified fact. Sources 1. Three breaches, endless leaks: Your AI supply chain needs canaries https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens dragos.com 2. AI Prompt Canaries: 3 Provider Breaches, ZDR Ignored https://explainx.ai/blog/dragos-hd-moore-two-million-canary-prompts-three-provider-breaches-zdr-ignored-ai-routers-2026 explainx.ai 3. blue team cyber defense blueprint - Black Eagle Group https://blackeaglegroup.io/blueteam blackeaglegroup.io 4. AI Supply Chain Security: What Canary Tokens Found | Dragos - Dragos https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens?ref=explainx dragos.com