Skip to main content

Dragos Planted Two Million Canary Prompts and Found Attackers Reading Them at Three AI Providers

Dragos says fake secrets hidden in about two million prompts were used by attackers at three AI providers, some despite Zero Data Retention flags. OpenRouter has removed three providers. What to check before you trust a privacy setting.

KBY News Desk11 min read
Detailed view of fiber optic patch cables connecting to a blue patch panel in a data center.

The security firm Dragos says it sent about two million AI prompts, each carrying its own fake secrets, to widely used AI services between July and November 2026, and found evidence that attackers were reading and using the contents at three AI providers, none of which knew until Dragos told them. Some of the stolen requests had been marked Zero Data Retention, the setting that is supposed to mean a prompt is not stored. On Sunday 22 November, OpenRouter, a gateway that routes customers’ requests to a pool of outside providers, said it had taken three providers off its platform, removed 125 endpoints from Zero Data Retention eligibility and notified customers who may have been affected. OpenRouter says the testing showed no breach of OpenRouter itself. Dragos has not named its three providers and OpenRouter has not named the three it removed, and neither says whether they are the same three.

#A fake key is an alarm that only rings once

The method is simple, which is why it is convincing. In a post published on 19 November, Dragos’s chief research officer HD Moore describes seeding prompts with unique URLs and canary credentials: AWS keys, DNS names and links that no legitimate system would ever use. The AWS keys are real keys tied to empty accounts with full logging. A secret like that exists in exactly one place, so any attempt to use it means somebody, somewhere, has read the prompt that carried it. Dragos says it can tie each attempt back to the exact prompt, model, router and privacy setting involved.

It also built what Moore calls “something like traceroute for AI inference”. Each request includes an image URL pointing back to Dragos, so every system that fetches it reveals an address and a timestamp, and slightly broken requests make each router wrap the error from the hop behind it, which exposes the chain. Dragos says it has traced about two thousand routes, and several hundred of them included a system it could not identify.

#Three providers, one front door

The most serious finding, in Dragos’s words, was “attackers actively stealing data from three different hosting providers”. The earliest incident dates back to the middle of July. The biggest involved a couple of hundred AWS keys, each sent in exactly one request to one provider during a few short windows in late October. Every request sent in those windows was stolen. Dragos says some of the stolen keys came from requests the provider had rejected before any model ran, which it reads as the attacker copying requests at the provider’s front door rather than somewhere deeper in the inference stack. In the other two breaches, some of the stolen requests were flagged as Zero Data Retention. Moore’s line on that: “A privacy flag doesn’t help when the attacker is already inside.”

What happened next to the keys will be familiar to anyone who has watched a leaked cloud credential. Dragos says the attacker tested each key within hours in some cases, listed S3 buckets, searched Secrets Manager and Parameter Store, then tried to enable hosted models in AWS Bedrock, create Bedrock API keys and add their own IAM users to keep access, driving tens of thousands of AWS API calls. One adversary appears to have compromised two of the providers at once, because secrets sent to both were used together, almost immediately, from a single rented server. On 1 November that attacker added ARTEX, which Dragos describes as an LLM-driven penetration testing framework, though only to confirm that keys it already held still worked. Two of its user agents, artex-agent and artex-judge-r121, appear in no public indicator list Dragos could find. Kali Linux and ARTEX both showed up in the user agent that AWS records in CloudTrail, but Dragos says that self-identification was present in only a small fraction of the traffic.

We think the front-door reading is the most important detail for anyone who buys inference. It is Dragos’s interpretation of the pattern rather than a confirmed root cause, and none of the three providers has been identified, but it suggests the interception happened before the point where retention settings are applied. If so, a retention flag set by the customer could not have protected those requests.

#The flag binds one company, and the prompt travels further

The other two findings needed no break-in. Dragos says sellers on an open inference marketplace listed models at the lowest price, switched on request logging and collected what flowed through. It describes individuals on what looked like home internet connections who opened the planted links and guessed at login pages by hand. In one case, a person on a Mac on a home broadband line in France took the passwords from two prompts and tried to log in as root over SSH, typing each attempt. Dragos also found a router that was quietly forwarding some traffic to the marketplace, an undisclosed hop, and says the router’s operator told it the forwarding had stopped.

The third finding is about promises. Dragos says a Chinese AI lab, which it does not name, fetched the secret links from more than half of the prompts sent to its models between July and October, including roughly two-thirds of the requests marked as no data collection or Zero Data Retention. Prompts sent to US hosting providers that resell the lab’s models triggered the same scraper. Twenty-four days after one prompt was sent, its secret link was posted into a chat on Feishu, the workplace app known as Lark outside China, 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 the provider, so whoever shared it was reading from a stored copy of the full prompt. Dragos’s conclusion: “Zero Data Retention is supposed to mean something. In practice, these prompts were retained, shared with other companies, and read by people.” That is one firm’s experiment, run on its own prompts, and the people and companies involved have not been named or given a chance to answer in public, so we would treat the individual cases as claims by Dragos and the pattern as the point.

#What the providers have said so far

OpenRouter’s announcement lists three provider notices. Phala published a preliminary incident report dated 10 November. It says an attacker took control of an inference virtual machine on a new GPU server that was serving traffic without the company’s full production hardening, chaining a legacy internet-facing service with management interfaces that trusted loopback access and did not independently authenticate callers. Phala says requests to DeepSeek V4 Flash 0731 that reached it through the gateway between 21 October 2026 at 00:00 UTC and 8 November at 05:35 UTC should be treated as exposed, and that every key, token or password in them should be rotated. It says canary evidence confirmed misuse only for requests sent on 21 and 22 October, but that the testing sampled traffic, so it keeps the wider window. Containment was complete on 8 November, and Phala says a full root cause analysis will follow.

Together AI published a notice on 21 November. It says prompt samples collected for routine testing, from requests to MiniMax M3, GLM-5.2 and Kimi K2.6 made mainly on 24 June and 4, 8 and 14 July, were kept in a private Hugging Face repository. An AI coding agent committed a credential for that repository to Together’s public GitHub repository, and an unauthorised party used it to download the samples, then searched them for credentials to other services. Five of the credentials tested were Dragos canaries. Together says it has found no evidence of further prompt leakage, is still reviewing what personal or customer data was in the samples, and has made Zero Data Retention the default for new accounts. The third notice on OpenRouter’s list is a privacy notice from Novita about MiniMax requests, which The Hacker News describes as a routing configuration issue that sent some requests on to MiniMax’s own API. Neither Phala, Together nor Novita is confirmed by either Dragos or OpenRouter as one of the three breached or removed providers, and OpenRouter says not all affected providers have posted publicly.

#Where we land

Our position is that this is a bigger deal than the usual “vendor had an incident” story, because it is about a control people have been told to rely on. Zero Data Retention is a promise made by the company that receives the request. It says nothing about the router in front of it, the reseller behind it, or an intruder sitting upstream of the retention logic. Any business that put Zero Data Retention in a data protection impact assessment as the reason prompts were safe should reread that sentence this week.

The pattern is also now hard to miss. Last week we wrote about AI agents hunting exposed AI apps for the provider keys stored inside them, and on 18 November Microsoft’s agent framework package on PyPI was poisoned to steal cloud credentials. Different routes, same prize: keys, tokens and passwords, which AI tooling gathers into prompts, context windows and gateways. Together’s account of an AI coding agent committing a credential is the same lesson from the supplier’s side.

For UK organisations, our reading is a practical one. If a prompt can contain personal data about staff, customers or patients, then every provider that handles it is part of your supply chain, including the ones behind a router that you never chose. You should be able to name them. That is a question for your data protection lead, and we are not offering legal advice here. For small businesses that do not run any of this directly, the equivalent is to ask whoever supplies your AI-enabled software, in so many words: “Which model providers does this product send our data to, through which router, and what do you do if one of them has an incident?” We would not accept “we use a secure provider” as an answer.

#What to do this week

  1. List where prompts actually go. Include the direct vendor, any router or gateway, any reseller marketplace and any agent framework configured with a router key. If you cannot draw the path, treat that as the finding.
  2. If you use OpenRouter, look for its notice to customers, and check whether any requests went to DeepSeek V4 Flash 0731 between 21 October 00:00 UTC and 8 November 05:35 UTC. Phala’s instruction for those requests is to rotate every secret they contained.
  3. If you used MiniMax M3, GLM-5.2 or Kimi K2.6 on Together AI’s serverless inference in June or July, watch for Together’s direct contact, and ask whether your prompts were in the stored samples.
  4. Rotate any credential that has ever been pasted into a prompt, and rotate the credential itself, not just the login that created it. On AWS, a console password reset does not revoke access keys, and our configuration trap on the subject covers how to find and deactivate them.
  5. Search CloudTrail for the user agents artex-agent and artex-judge-r121, and for Kali Linux. Dragos says these showed up, though only in a small fraction of traffic, so a clean result proves little.
  6. Pick upstream providers yourself instead of accepting a router’s default pool, and verify that a privacy setting really changes where requests go. Dragos’s advice is to stay away from reseller marketplaces.
  7. Plant a tripwire. Dragos suggests adding a fake credential, for example a free Thinkst Canarytoken, to the context your agents already send, and treating any use of it as an incident.

None of that means stopping use of AI services, and Dragos says so itself. It does mean deciding where prompts go rather than leaving the cheapest route to decide.

#What we expect, and what would change our mind

We expect more provider notices. Dragos says it is working with about a dozen providers, OpenRouter says it will keep its list updated, and Dragos reports that planted keys and links were still being used months after the first thefts. We expect canary-style testing to become a standing check at gateways: OpenRouter says it is already adapting the researchers’ approach so it can run on an ongoing basis. We would be wrong about the first if the list stays at three, and the unnamed breached and removed providers turn out to be the same three, with nothing further found.

The question we would most like answered is who the mid-July provider is. Dragos says it first blamed an internal misconfiguration, then a partner’s data processing, before its own investigation found that attackers had compromised its internal credentials, and that it had still not publicly disclosed the incident when Dragos published on 19 November. If you use a gateway, that is a fair question to put to it directly.

Evidence trail

Sources and verification

Primary documentation and external technical references used in this article.

  1. 01Dragos, "Three breaches, endless leaks: Your AI supply chain needs canaries" (HD Moore, 19 November 2026)dragos.com
  2. 02OpenRouter, "Data Processing Incidents on Inference Providers" (22 November 2026)openrouter.ai
  3. 03Phala, "November Preliminary Incident Report" (10 November 2026)phala.com
  4. 04Together AI, "Security event: Prompt leakage for requests in June and July of 2026" (21 November 2026)together.ai
  5. 05The Hacker News, "OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms"thehackernews.com

About this page

Credited to

Dates

Published

Corrections

No corrections are listed for this page. Some review sweeps corrected pages without naming each one; those are in the corrections log.

Reader Interaction

Comments

Add a thoughtful note on Dragos Planted Two Million Canary Prompts and Found Attackers Reading Them at Three AI Providers. Comments are checked for spam and held for moderation before appearing.

Loading comments...

Get new fixes by email.

New runbooks, configuration traps and fixes from KBY Technologies, about twice a week.