OpenRouter AI inference provider breach: Dragos canaries caught months of prompt theft, stolen AWS keys abused; OpenRouter drops 3 providers
SockAnalyst · Security+ certified analyst.
What happened
On November 22, 2026, OpenRouter said it has off-boarded three inference providers, removed 125 endpoints from Zero Data Retention (ZDR) eligibility (flagged providers plus precautionary removals) and notified all potentially affected customers, after security researchers reported that some providers had data compromised or were not honoring data policies. OpenRouter says the findings showed no breach of OpenRouter itself; it has not named the three providers. The research is Dragos's (chief research officer HD Moore, published November 19): from July through November 2026 Dragos sent about two million prompts carrying unique canary URLs and credentials (including real AWS keys tied to empty, fully logged accounts) and watched where they surfaced. Findings: (1) attackers were stealing data inside at least three hosting providers, the earliest from mid-July, and none of the three knew until Dragos told them. In the largest case every request Dragos sent in a few late-October windows was stolen, including requests the provider rejected before a model ran, so the copying happened at the provider's front door. Stolen AWS keys were tested within hours, then used to list S3 buckets, search Secrets Manager and Parameter Store, try to enable AWS Bedrock models and create Bedrock API keys, and add IAM users (tens of thousands of API calls). One adversary appeared to hold two providers at once; its tooling moved from a fake Chrome-on-Windows script to Boto3 in OrbStack to the AWS CLI on Kali Linux, and on November 1 it used the ARTEX AI pentest framework, whose user agents artex-agent and artex-judge-r121 showed up in CloudTrail. In two breaches some stolen requests were flagged ZDR. (2) Sellers on an AI inference marketplace listed models at the lowest price, enabled request logging and used the captured secrets (URL logins, hand-typed root SSH attempts, AWS key tests). (3) Privacy settings were ignored: a Chinese AI lab's systems fetched secret URLs from more than half the prompts sent to its models, including about two-thirds of no-data-collection/ZDR requests, also via U.S. resellers; one seeded link turned up in a Feishu chat 24 days later. Provider disclosures linked by OpenRouter: Phala (Nov 10 preliminary report) says an attacker took control of an inference VM and could see in-flight request data; treat DeepSeek V4 Flash 0731 requests routed to Phala by OpenRouter between Oct 21 00:00 UTC and Nov 8 05:35 UTC as exposed and rotate any secrets in them. Together AI says an unauthorized party downloaded stored prompt samples (MiniMax M3, GLM-5.2, Kimi K2.6 requests from June-July) after an AI coding agent committed a credential to its public GitHub repository. Novita says a routing configuration issue sent some MiniMax requests on to MiniMax's official API. Neither Dragos nor OpenRouter has said whether those are the same three providers.
Who's affected
Any organization or developer whose apps, AI agents or coding assistants send prompts through AI routers, marketplaces or third-party inference providers, especially if prompts contained credentials, API keys, internal URLs, source code or customer data. OpenRouter users should check for its notification; Phala, Together AI and Novita users should follow those notices. Cloud accounts whose keys ever appeared in a prompt are at direct risk.
✅ Fix / patch
- If you got an OpenRouter, Phala, Together AI or Novita notice, treat the listed prompt windows as exposed and rotate every key, token and password that appeared in those prompts (Phala: DeepSeek V4 Flash 0731 via OpenRouter, Oct 21-Nov 8, 2026).
- Assume any secret ever pasted into a prompt or read by an AI agent may be exposed. Find where agents and pipelines send environment variables, config files or repo contents to LLMs and rotate those credentials.
- Pin the upstream providers you trust in your router settings instead of the default lowest-price pool, avoid reseller marketplaces for sensitive work, and verify that privacy flags (ZDR, no data collection) actually change where requests go.
- Turn on prompt secret-scrubbing where your gateway offers it, and seed a canary token (for example a free Thinkst Canarytoken) into agent context so any use of it is treated as an incident.
- In AWS, alert on the attacker chain Dragos observed: new IAM users or access keys, Bedrock model enablement or Bedrock API key creation, Secrets Manager/Parameter Store enumeration, and CLI user agents from Kali or ARTEX.
🔎 Quick hunt check
Dragos published two tool user agents seen in CloudTrail, artex-agent and artex-judge-r121, plus the attacker's behavior; it published no IPs or domains. Queries: (1) AWS CloudTrail events whose user agent contains artex-agent, artex-judge or kali; (2) any access key that, within 24 hours, both enumerates data (ListBuckets, ListSecrets, DescribeParameters/GetParametersByPath) and tries persistence or Bedrock abuse (CreateUser, CreateAccessKey, CreateServiceSpecificCredential, PutUseCaseForModelAccess, CreateFoundationModelAgreement).
MITRE ATT&CK: T1078.004, T1619, T1555.006, T1136.003, T1098.001
IOCs
- Files / paths:
AWS CloudTrail user agents: artex-agent, artex-judge-r121 (ARTEX AI pentest framework, per Dragos),Kali Linux in the AWS CLI user agent (tooling seen by Dragos; also used by legitimate testers)
Dragos published no IPs, domains or hashes, and masked attacker addresses in its figures. The user agents appeared in only a small fraction of attacker traffic, so the behavior query matters more than the strings. The 125 endpoints OpenRouter removed from ZDR are not 125 confirmed breaches.
Hunt queries
Starting points built only from published indicators. Field names vary; tune before use.
```OpenRouter AI inference provider breach: CloudTrail ARTEX/Kali user agents + Dragos-observed recon-then-persistence/Bedrock chain by one access key within 24h```
index=* sourcetype=aws:cloudtrail earliest=07/01/2026:00:00:00
| eval ua=lower(userAgent), key=coalesce('userIdentity.accessKeyId', "unknown"), arn='userIdentity.arn'
| eval recon=if(match(eventName, "^(ListBuckets|ListSecrets|GetSecretValue|DescribeParameters|GetParametersByPath)$"), 1, 0), persist=if(match(eventName, "^(CreateUser|CreateAccessKey|CreateLoginProfile|AttachUserPolicy|CreateServiceSpecificCredential|PutUseCaseForModelAccess|CreateFoundationModelAgreement|PutFoundationModelEntitlement)$"), 1, 0), tool_ua=if(match(ua, "artex-agent|artex-judge|kali"), 1, 0)
| where recon=1 OR persist=1 OR tool_ua=1
| bin _time span=24h
| stats sum(recon) as recon_calls sum(persist) as persist_calls sum(tool_ua) as tool_ua_calls values(eventName) as events values(userAgent) as user_agents values(sourceIPAddress) as src_ips by _time, key, arn
| where tool_ua_calls>0 OR (recon_calls>0 AND persist_calls>0)
| sort - tool_ua_calls - persist_calls// OpenRouter AI inference provider breach: CloudTrail ARTEX/Kali user agents + Dragos-observed recon-then-persistence/Bedrock chain by one access key within 24h
let Since = datetime(2026-07-01);
let ReconEvents = dynamic(["ListBuckets", "ListSecrets", "GetSecretValue", "DescribeParameters", "GetParametersByPath"]);
let PersistEvents = dynamic(["CreateUser", "CreateAccessKey", "CreateLoginProfile", "AttachUserPolicy", "CreateServiceSpecificCredential", "PutUseCaseForModelAccess", "CreateFoundationModelAgreement", "PutFoundationModelEntitlement"]);
AWSCloudTrail
| where TimeGenerated >= Since
| extend ToolUA = UserAgent has_any ("artex-agent", "artex-judge", "kali"), Recon = EventName in (ReconEvents), Persist = EventName in (PersistEvents)
| where ToolUA or Recon or Persist
| summarize ToolUACalls = countif(ToolUA), ReconCalls = countif(Recon), PersistCalls = countif(Persist), Events = make_set(EventName, 30), UserAgents = make_set(UserAgent, 10), SourceIps = make_set(SourceIpAddress, 20), FirstSeen = min(TimeGenerated) by UserIdentityAccessKeyId, UserIdentityArn, Day = bin(TimeGenerated, 24h)
| where ToolUACalls > 0 or (ReconCalls > 0 and PersistCalls > 0)
| order by ToolUACalls desc, PersistCalls desc⏱ CISA deadline
Not in CISA KEV as of Nov 22, 2026.
Sources
- OpenRouter: Data Processing Incidents on Inference Providers (Nov 22, 2026)
- Dragos (HD Moore): AI Supply Chain Security: What Canary Tokens Found (Nov 19, 2026)
- Phala: November Preliminary Incident Report (Nov 10, 2026)
- Together AI: Security event: prompt leakage for requests in June and July of 2026
- Novita: Privacy notice regarding MiniMax requests
- The Hacker News: OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms (Nov 22, 2026)
- Ars Technica: Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced. (Nov 2026)
Related playbooks
- Microsoft Entra Connect Health API flaw exploited: admin email addresses and tenant metadata exposed, Microsoft notifying affected tenants · Breach · 2026-11-22
- Barracuda cloud breach: unrevoked AI-vendor access key let an ARTEX-style intruder roam 10 cloud accounts for 44 hours · Breach · 2026-11-16
- Northflank breached: assume every env var was decrypted · Breach · 2026-10-30
Get the SockAnalyst Weekly Brief free, every Thursday.
One email a week with the KEV deadlines, patches and hunt queries that matter. New playbooks still post here on the site every morning.
Subscribe free
SockAnalyst