Dragos Seeded 2 Million Prompts With Canaries; Amazon's Decoys Caught ARTEX Agents Grading Stolen AI Keys
Security / analysis
Dragos Seeded 2 Million Prompts With Canaries; Amazon's Decoys Caught ARTEX Agents Grading Stolen AI Keys
Two reports from November 13 and 19 show AI credentials being harvested and spent by machines, and a Chinese lab fetching secrets from ZDR prompts.

Two reports published within a week of each other describe the same market from opposite ends: Amazon watched attackers' AI agents work through exposed AI gateways, and Dragos fed about two million canary-laden prompts to AI services and watched where the secrets surfaced. Dragos found data from at least three AI providers exposed, and says "none of the three providers knew about their breach until we notified them." Amazon assesses, with moderate confidence, that the agents ran on ARTEX, an open-source AI penetration-testing framework.
Neither report establishes that one actor is behind both. Together they show stolen AI credentials being collected, graded and spent by machines, and show that a provider's privacy flag does not stop the data being read downstream.
What Dragos seeded and what came back
HD Moore, Chief Research Officer at Dragos, wrote on November 19 that his team sent "approximately two million prompts" to widely used AI services between July and November 2026. Each carried unique URLs and canary credentials: first Thinkst Canarytokens, then a purpose-built service that derives a distinct hostname, URL and credential per prompt and answers on DNS, web, git, SSH, databases and mail. The AWS keys were tied to empty, fully logged accounts.
The post reports three findings:
- Provider compromises. The earliest incident dates to mid-July. The largest involved "a couple hundred" AWS keys, each sent in one request. One key seeded on October 21 was first used the next day and was still in use three weeks later, and the stolen keys drove "tens of thousands of AWS API calls" against S3, Secrets Manager, Parameter Store, Bedrock and IAM.
- Reseller exposure. Sellers on an AI inference marketplace could read and keep customer prompts through ordinary logging. One seller added about a thousand hidden tokens to every prompt.
- Retention that ignored settings. A Chinese AI lab fetched secret URLs from more than half of the prompts sent to its models between July and October, about two-thirds of those flagged for no data collection. A link from a zero-data-retention request to a U.S. provider was posted in a Feishu chat 24 days later and opened from three desktops and three phones in Beijing and Shanghai within two minutes.
The post does not name the three providers, and every figure in it is Dragos's own, from prompts Dragos wrote and accounts it controls; no outside party has reproduced them in anything I fetched. It traces about 2,000 routes and says notification of about a dozen providers began in early November; one has published a preliminary notice and one has not disclosed.
What Amazon's decoys recorded
CJ Moses, CISO of Amazon Integrated Security, wrote on November 13 that from October 3 to November 5, 2026 Amazon's MadPot decoy sensors, configured to look like LLM inference APIs, received requests from AI agents that carried the agents' own campaign memory. The most prominent campaign was theft of credentials for AI models, run so that stolen keys paid for the inference powering more hunting.
The standing instructions Amazon quotes are operational: work through a fixed sequence of countries, prioritise enterprise production sites over personal projects, grade every stolen credential by whether it worked and what it could reach, and treat credentials exposed publicly for more than 30 days as low-value inventory. Amazon also assesses, again with moderate confidence, that the operators were separate Chinese-nexus groups, citing Simplified Chinese prompts and China time zone settings. It says it could not observe activity on real systems and cannot verify the outcomes the agents reported.
The Hacker News, summarising the Amazon post on November 16, adds that ARTEX's developer took its GitHub repository down on October 8 after abuse reports, though forks, archived copies and a commercial version called ARTEX PRO remain. It lists the targets: One API, New API, LiteLLM, Dify, Langflow, RAGFlow, Open WebUI and others.
Where the two overlap
| Observation | Dragos | Amazon |
|---|---|---|
| Method | About 2 million canary prompts, July to November | Decoy LLM APIs, October 3 to November 5 |
| ARTEX | Added to attacker tooling on November 1; user agents artex-agent and artex-judge-r121 | Moderate-confidence attribution of the campaigns |
| What attackers wanted | AWS keys, then Bedrock access and new IAM users | Credentials for AI models and cloud AI services |
| Limits | Providers unnamed | Cannot verify reported outcomes |
Dragos lists ARTEX among the tooling it saw against its canary accounts, after earlier traffic from a Chrome-impersonating script, Boto3 running in OrbStack and the AWS CLI on Kali Linux. A new framework appearing after scripts and manual tools fits Amazon's description of agents doing the grading, but neither report says the same operators are involved.
What to do with this
Rotate any key that has ever been pasted into a prompt, because Dragos's canaries show prompts are not a private channel even when a retention setting says so. Finish gateway installs and change default master keys: The Hacker News's account of the Amazon post lists LiteLLM's default master key among the ARTEX playbook items, and the product has already been in the news for a CISA listing with conflicting endpoint details. Agent tooling is also spreading the other way, as in Claude-BugHunter's skill bundle, which defenders and attackers can both run.
Dragos itself suggests seeding your own canaries, going directly to first-tier providers rather than reseller marketplaces, and enabling secret filtering, while noting it is not a complete fix.
Sources
More in Security
- 01GPGME's CVE-2026-129975 Double Free Needs Only a Crafted Key, and 2.3.1 Is the FixGnuPG disclosed five days early because GitLab had to update, and no CVSS score exists yet to rank it.
- 02REDCap's CVE-2026-90817 Needs Only a Public Survey Hash, and CISA Listed It 59 Days After the CVEVanderbilt's survey platform is fixed in 16.0.49 LTS, 17.3.10 LTS and 17.4.4, and Safe Security reports a miner installed with cron persistence and an SSH key.
- 03LiteLLM's CVE-2026-40217 Is on CISA's KEV List, but Advisory and Google Name Different EndpointsGitLab's advisory says the bug needs a proxy-admin credential, while Google's report says attackers used it to turn standard virtual keys into admin ones.
- 04Ghostscript's Upload2Shell Flaw, CVE-2026-125396, Escapes -dSAFER and Is Fixed in 10.08.0The flaw is scored 7.8 to 8.5 as a local bug, yet researchers confirmed command execution through CUPS print queues, Roundcube message view and WordPress uploads.