Amazon's agents paid with stolen keys. Storm-3358 is not this hunt.
FBI FLASH-20261113-001, dated 13 November, says ARTEX has been used since August and some intrusions reached privileged access in under three hours. Amazon's MadPot decoys, 3 October–5 November, captured the agents' own memory — hunt internet-exposed LiteLLM, Dify, One API and other AI apps, steal provider keys, feed working keys back into the model gateway. Amazon cannot verify the outcomes. Changing the default master key is not the inventory.
Bottom line up front
CONFIRMED On 13 November 2026 the FBI published FLASH-20261113-001 (TLP:CLEAR), a threat overview of the open-source agentic pentest framework ARTEX. The bureau’s sentence, as restated in the FLASH and in Friday’s secondary reporting: a variety of malicious actors have used it since at least August 2026 across financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications and legal; some intrusions went from initial targeting to privileged access in under three hours. The same day, Amazon Threat Intelligence (CJ Moses) published what MadPot actually captured: from 3 October to 5 November, ARTEX agents sent their own campaign memory to decoys posing as LLM inference APIs. The lead campaign in those records is industrial-scale theft of AI-provider keys from internet-exposed gateways and app builders, with stolen keys fed back into the framework’s model gateway to pay for the next hunt.
Do not collapse five facts.
- This is not [Storm-3358](/posts/2026-11-08-storm-3358-artex-financial/), and it is not Storm-0940. The 8 November Daily Top was Microsoft’s financially motivated cluster: anonymous upload and SSRF, then WISP / vshell / AdaptixC2. Storm-0940 was the [ConnectWise Cloud Backup publisher-secret path](/posts/2026-10-17-connectwise-skykick-m365-oauth/). The FLASH recites both as examples. Amazon’s MadPot records describe a different door: default admin credentials and unauthenticated management on LiteLLM, One API, Dify, Langflow, RAGFlow, Open WebUI, Ollama and related products. Closing last month’s seven Application IDs, or hunting WISP C2, does not inventory an internet-facing LiteLLM master key.
- Amazon did not watch the victims. MadPot saw prompts, tool histories, knowledge-graph state and reported findings that agentic frameworks send to an LLM. Amazon: it did not observe that activity as it happened and cannot verify the outcomes the agents reported. AWS customers who may have been affected were notified. Treat agent self-reports as ASSESSED targeting, not a confirmed census of stolen keys.
- The FLASH is a class warning, not a new CVE. Live catalog this run is still
2026.11.13/ 1781,dateReleased2026-11-13T20:24:18.3008Z. No overnight KEV add. ARTEX has no CVE and no KEV row. The developer pulled the original GitHub repo on 8 October; the FBI says forks, archives, localized copies and a commercial ARTEX PRO remain. Early November: researchers counted more than 1,600 publicly exposed ARTEX consoles (FBI); Censys’s 4 November count in secondary reporting was 1,465. Those are console counts, not confirmed intrusions. - This is not the October LiteLLM wheel. [LiteLLM 1.107.0](/posts/2026-10-25-litellm-11070-cook-pth/) was
cook.pthin a PyPI/Docker tag. Amazon’s playbook is an internet-exposed management plane and a default master key. Uninstalling a wheel does not close an admin UI that still answers on 0.0.0.0. - Monday’s KEV clocks are a different ticket. [CVE-2026-123039](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/) (Check Point Gaia DynamicCLI) and CVE-2026-114713 (UnrealIRCd labeled-response) are due today, 16 November, forensic triage Yes, ransomware use Unknown. Yesterday’s Daily Top was [Cursor’s Fireworks-exposed vendor key](/posts/2026-11-15-cursor-fireworks-claude-code-marketplace/). None of those grew overnight. A Jumbo Take is not this gateway hunt.
The operator sentence: **if you run LiteLLM, Dify, One API, Langflow, RAGFlow, Open WebUI, Ollama or any other aggregator of provider keys on an address the internet can reach, the floor today is take the admin plane off the public internet, rotate every stored provider and cloud key from a clean host, and hunt Cloudflare Workers named artex-*. A missing FBI email is not a green row. Monday’s Check Point Jumbo is a different ticket.**
What happened
16 November 2026, after yesterday’s 06:00 ET Daily Top. Yesterday’s live catalog was 2026.11.13 / 1781. It still is. CISA added nothing overnight. Yesterday’s Daily Top was Cursor’s 13 November investigation of a Fireworks-exposed vendor key into a marketplace cache. The FBI FLASH and Amazon’s MadPot post are both dated 13 November; the increment for this run is Friday afternoon secondary (BleepingComputer, 15 November 13:29 ET — after yesterday’s 06:00 ET run) plus Monday’s The Hacker News recap of Amazon, which publicly tied the FLASH to a product-level AI-gateway hunt that was not Friday’s Cursor laptop story.
Amazon’s picture, paraphrased from the 13 November AWS Security Blog, not a reproduction:
- MadPot sensors presented popular LLM-compatible chat-completions interfaces. Agentic frameworks treated them as inference backends. Because those frameworks retain conversation history, individual requests often carried the full operational history of a campaign — activity MadPot does not monitor.
- Amazon assesses moderate confidence that the campaigns ran on ARTEX (prompt references and other technical artifacts). It also assesses moderate confidence that the operators were disparate Chinese-nexus groups (Simplified Chinese in prompts, China time-zone settings, targeting consistent with Chinese interests, tooling historically associated with that nexus). Sample standing instructions: exclude mainland China targets and government domains; send all outbound traffic through proxies; work a fixed country sequence; prioritize enterprise production over personal projects; grade credentials L1/L2/L3; treat keys exposed publicly for more than 30 days as low-value inventory.
- Campaign 1 — AI credentials. Find internet-exposed AI apps that aggregate provider keys. Extract the upstream keys. Validate them. Store them in a managed pool. Feed working keys into the framework’s own model gateway so stolen inference pays for the next hunt. Named targets: gateways and proxies One API, New API, Veloera, Sub2API, LiteLLM; app builders Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama. Playbook entries Amazon names at a high level: uninitialized installs an outsider can claim; anonymous management and provider-configuration endpoints; default administrator credentials such as the LiteLLM default master key; CVE-2025-3248 (Langflow); CVE-2026-28797 (RAGFlow SSTI); an
isCliRequestheader authentication bypass in One API-family systems; SSRF through URL-fetch features toward cloud metadata; secrets in client-side JavaScript, source maps, debug routes and environment files. Discovery also drew on internet scanners, certificate transparency, GitHub (secrets outside automated revocation), npm and PyPI tarballs, container image layers, notebooks, model-hosting spaces, paste sites and Telegram. - Campaign 2 — KYC deepfakes. Agents tried live deepfake video and forged ID photos against two cryptocurrency exchanges. Amazon: the vendor’s server-side liveness verdict rejected every recorded attempt. Document-check success is not stated.
Cloudflare’s 13 November network note is a separate observation, not Amazon’s victim list. Cloudflare: a few weeks earlier ARTEX was the only agentic pentest orchestrator it tracked; in the three days before the post, more than 50 framework families attempted to scan Cloudflare customers; daily requests from AI pentest/exploitation tools rose more than 30× since 1 October; as of 12 November, ARTEX and forks were over 60% of known orchestrators it tracks; the ecosystem around the pulled repo is over 12,000 repositories with 204 public forks that carry their own code changes. Cloudforce One IR, last eight weeks: static, never-expiring account-level API tokens in CI/CD, then a pivot into the customer’s Cloudflare account. In the ARTEX case it examined: new Worker scripts named artex-, artex-pgrecon-, artex-pgw-, artex-do-; more than 69 requests per minute; 400 database-dump requests in an hour. Whether those dumps returned data is not disclosed. How the token was first stolen is not disclosed.
The FLASH’s notable campaigns, as the FBI wrote them, are already in the archive: South Korean financial institutions, personal information of approximately 67,000 people; ConnectWise Cloud Backup enabling access toward more than 2,250 downstream tenants. Do not treat the FLASH as a new ConnectWise wave. The Hacker News, 16 November: the alert summarizes 11 earlier reports and describes no new incident. The new defender work is Amazon’s product playbook plus Cloudflare’s Worker/token hunt.
Not this increment as a reprint. [Yesterday’s Cursor / Fireworks marketplace](/posts/2026-11-15-cursor-fireworks-claude-code-marketplace/) is a VSIX cache on developer laptops. [Check Point CVE-2026-123039](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/) is due today — catalog unchanged. UnrealIRCd CVE-2026-114713 is the other Monday clock, still IRC. NoisyTCP CVE-2026-99288 (BleepingComputer, 16 November 04:30 ET) is an 11-year-old Linux TCP race that panics the kernel; the researcher and Red Hat describe DoS, not code execution, and it is not on KEV this run. Kaspersky’s BO Team / Log4Shell vCenter IR is a 2025 incident recapped today — not this gateway.
Why it matters
Three operational reasons this is the daily.
First: the false close is “we handled ARTEX.” Three Microsoft posts, one FBI FLASH, one Amazon decoy paper, one Cloudflare IR note. Five doors. Storm-0940 was a trusted app in the tenant. Storm-3358 was upload/SSRF then a RAT. GitHub-secret hunting was OAuth client secrets. Amazon’s records are AI apps that hold other people’s keys. Cloudflare’s case is a static CI/CD token with no IP restriction. One change ticket labeled ARTEX will miss four of five.
Second: an AI gateway is a credential vault. LiteLLM, One API and Dify exist to multiplex provider keys. One weak admin page exposes many downstream accounts. Amazon’s own sentence: none of the techniques were new; what agents change is speed and persistence. Default master keys, unfinished installs, and unauthenticated /v1 endpoints are the same class of miss that put RMM consoles on this site last week. They are not a novel parser bug.
Third: the self-funding loop is a billing and identity problem, not a malware signature. If stolen keys pay for the next inference, the first telemetry may be unexpected spend or unusual model usage, not a hash. That is still IT/cloud. This is not confirmed OT or ICS process impact. A plant whose engineers left Open WebUI on a VPS has an IT incident until someone shows the same host writing a controller.
Who / what is affected
| Asset | What the sources say | Floor / action | Who should care first |
|---|---|---|---|
| Internet-exposed LiteLLM, One API / New API / Veloera / Sub2API, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama | Amazon MadPot records: playbook of defaults, unauth management, named CVEs, key extraction | Take admin/inference off the public internet; change defaults; rotate every stored provider and cloud key from a clean host | AI platform owners, CISO for coding/inference tools |
| Langflow before 1.3.0 | Amazon names CVE-2025-3248; already on KEV (5 May 2025), ransomware use Known | Patch; assume prior internet exposure is a compromise check | AppSec / vuln mgmt |
| RAGFlow 0.24.0 and earlier (NVD) | Amazon names CVE-2026-28797 SSTI; advisory: registered account, self-registration on by default | Upgrade / sandbox templates; disable open signup | Same |
| Cloudflare account with account-level API tokens used in CI/CD, no expiry / no IP allow-list | Cloudflare IR: Workers artex- / artex-pgrecon- / artex-pgw- / artex-do-; 69+ rpm; 400 dump requests/hour | Hunt those names; rotate tokens; add expiry and IP restriction | Cloudflare / DevOps |
| AWS customers Amazon assessed as potentially referenced | Amazon: notified as appropriate | Follow the vendor notice; do not wait for a public victim list | AWS account owners |
| Exposed ARTEX consoles (~1,465–1,600+ public) | FBI / Censys counts | If you left a console on the internet, treat it as an unauthenticated pentest kit with your LLM keys in it | Anyone who cloned ARTEX “for testing” |
| Check Point Gaia Mobile Access; UnrealIRCd 6 linked nets | KEV due today | [Friday’s Jumbo/hunt](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/); UnrealIRCd 6.2.7.1 | Perimeter / IRC — not this gateway |
Amazon did not publish a confirmed victim count. “Hundreds of targets” is what the agent records described. Do not brief the board that Amazon confirmed hundreds of breaches.
Technical context
High-level, defensive only. Not a reproduction.
ARTEX, as the FBI and vendors describe it, is a Go agentic pentest framework. An operator gives a target and a goal. A planner splits work for worker agents. Those agents call external LLMs (the FLASH and BleepingComputer list Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok, Kimi among supported backends) and conventional tools. Amazon: planner, worker, verification and compression agents sharing a persistent knowledge graph. The model does not invent the CWE. It retries T1190 Exploit Public-Facing Application and T1552 Unsecured Credentials until something answers.
The LiteLLM-shaped miss is identity, not a new CVE this run: a default master key that is also the admin login (and, in prior public research not restated here as a how-to, has been described as the JWT signing secret). Amazon named the default master key as a playbook item. Rotate it. Do not leave a weekend VPS install unfinished.
Langflow CVE-2025-3248 is missing authentication leading to RCE in versions before 1.3.0; it has been on KEV since May 2025. RAGFlow CVE-2026-28797 is SSTI as a logged-in user with self-registration on by default. One API-family isCliRequest is an auth bypass Amazon named without versions or a fix. Those are patch and exposure tickets. They are not a walkthrough.
ATT&CK mapping for defender queries, not a playbook: T1190 Exploit Public-Facing Application, T1552 Unsecured Credentials, T1078 Valid Accounts (stolen keys / default admin), T1526 Cloud Service Discovery, T1098 Account Manipulation (Cloudflare Workers / persistence attempts), T1606.002 Forge Web Credentials (session-token validation in the Cloudflare case), T1530 Data from Cloud Storage Object.
Public PoCs, forks, and ARTEX PRO installers are untrusted input. They are not required to contain.
Threat / exploitation status
| Claim | Status | Confidence |
|---|---|---|
| FBI FLASH-20261113-001, 13 Nov, TLP:CLEAR, ARTEX used since at least Aug 2026, some intrusions to privileged access in under three hours | CONFIRMED as FBI statement | High |
| FLASH recites ConnectWise 2,250+ downstream tenants and SK financials ~67,000 people | CONFIRMED as FBI summary of earlier reporting | High for the recitation; those campaigns were already published |
| FLASH describes a new incident beyond those recaps | No — THN 16 Nov: summarizes 11 earlier reports, no new incident | High they are recaps |
| Amazon MadPot captured ARTEX-agent campaign memory 3 Oct–5 Nov, including AI-key theft playbook against named products | CONFIRMED as Amazon observation of decoy traffic | High for what MadPot received |
| Stolen keys actually worked / funded further hunts / reached named orgs | ASSESSED from agent self-reports; Amazon cannot verify | Moderate at most — do not treat as confirmed theft |
| Operators are disparate Chinese-nexus groups | ASSESSED (Amazon, moderate confidence) | Moderate — language/timezone/targeting, not a named APT |
| Framework used is ARTEX | ASSESSED (Amazon, moderate confidence) | Moderate |
Cloudflare: 30× AI-pentest request growth since 1 Oct; 50+ families in three days; artex-* Workers in one IR case | CONFIRMED as Cloudflare network/IR | High for Cloudflare’s own telemetry; UNKNOWN whether that case is Amazon’s campaign |
| Same operation as Storm-3358, Storm-0940, Cursor marketplace, Check Point 123039 | No evidence they are one operation | High they are separate queues |
| CISA KEV add overnight; CVE for ARTEX | No — catalog still 2026.11.13 / 1781 | High |
| OT/ICS process impact | UNKNOWN / not claimed | — |
| Ransomware use of this key-theft campaign | UNKNOWN / not claimed | — |
What defenders should do
- Inventory the AI apps before you trust a missing notice. Which LiteLLM / One API / Dify / Langflow / RAGFlow / Open WebUI / Ollama / FastGPT / LobeChat instances exist, including weekend lab boxes. Whether the admin UI or unauthenticated provider-settings endpoint answers on the public internet. Whether the LiteLLM default master key (or equivalent) was ever changed. Whether an install was left unfinished. Amazon notified some AWS customers. Silence is not a green row for a VPS you stood up in October.
- Rotate from a clean host, then hunt spend. Every provider key (OpenAI, Anthropic, others) and cloud credential those apps stored. Then the cloud identities those keys opened — managed model services, secrets/parameter stores, application identities. Amazon: agents prioritized premium frontier models over general cloud permissions. Watch unexpected inference spend and unusual model usage. Pull GitHub, npm, PyPI, container layers and paste sites for the same secrets; Amazon says the agents did.
- Hunt Cloudflare and CI/CD as if the token never expired. Worker scripts named
artex-,artex-pgrecon-,artex-pgw-,artex-do-. Account-level API tokens with no expiry and no IP allow-list. 400 dump-shaped requests in an hour is a volume tell, not a hash. Replace static CI/CD tokens with short-lived, IP-restricted, least-privilege credentials. Cloudflare: new tokens carry acfut_prefix that secret scanners can detect.
- Patch the two named CVEs and keep the Monday clocks. Langflow CVE-2025-3248 → 1.3.0 or later. RAGFlow CVE-2026-28797 → current release that sandboxes templates; disable open self-registration. [Check Point CVE-2026-123039](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/) Jumbo takes (R82.20 Take 14+, R82.10 Take 52+, R82 Take 132+, R81.20 Take 174+) and PHP2Shell on Mobile Access — due today, forensic triage Yes. UnrealIRCd CVE-2026-114713 → 6.2.7.1 or the 7 November hot-patch, or
blacklist-module "labeled-response"until then — due today. [SonicWall CVE-2026-19868](/posts/2026-11-07-sonicwall-cve-2026-19868-kev/) is due 20 November. Outlook CVE-2026-124194 and Chromium CVE-2026-125350 are due 24 November.
- Keep the ARTEX tickets separate in the queue. Storm-0940 / SkyKick: service-principal sign-ins. 5 November secrets: public GitHub OAuth. Storm-3358: web root and WISP/vshell. This run: AI-gateway defaults and stored provider keys, plus Cloudflare Workers. FBI: report ARTEX-associated activity to a field office or IC3. IP indicators in the FLASH and in Amazon’s table age out; Amazon’s eleven IPv4s (Campaign 1: 104.28.225[.]223, 31.22.91[.]156, 150.136.38[.]110, 128.242.115[.]131, 216.176.239[.]231; Campaign 2: 212.107.28[.]51, 103.142.140[.]158, 104.28.251[.]98, 104.28.251[.]99, 104.28.219[.]97, 125.103.213[.]41) are a start, not a blocklist. Secondary reporting points at Cisco Talos Snort rules 7:76–7:78 / 1:67494–1:67496 and a MalwareMustDie signature pack from Friday — use as supplementary, not as the close.
RWP assessment
Score 9.0. It outranks a second day on [Check Point CVE-2026-123039](/posts/2026-11-14-checkpoint-cve-2026-123039-kev/) (catalog unchanged; due-day clock already published Friday), UnrealIRCd 114713 (same catalog, IRC), NoisyTCP CVE-2026-99288 (DoS, no KEV, no observed exploitation this run), BO Team / Log4Shell vCenter (Kaspersky IR of a 2025 case, recapped today), and a reprint of [Cursor / Fireworks](/posts/2026-11-15-cursor-fireworks-claude-code-marketplace/).
The material development is not “ARTEX exists.” Microsoft named it in October. The 8 November Daily Top named Storm-3358. Overnight the FBI put a TLP:CLEAR FLASH on the class, and Amazon published a product-level playbook for internet-exposed AI gateways that is not WISP, not SkyKick, and not cook.pth. We will not invent a confirmed key-theft census, a named APT, a KEV row, ransomware, or OT process impact. We will not collapse Storm-3358, Storm-0940, Cursor’s marketplace cache, and this decoy paper into one victim list. Amazon’s moderate-confidence attribution stays ASSESSED.
What we are watching
- Whether Amazon, Cloudflare or a victim names a confirmed key that worked, a loss figure, or a CVE for the One API
isCliRequestbypass. - Independent IR that finds
artex-*Workers or LiteLLM default-key use outside Amazon’s notified AWS set. - Any CVE or KEV row for ARTEX itself — none this run.
- Check Point 123039 / UnrealIRCd 114713 through today’s BOD clock; SonicWall 19868 toward 20 November; Outlook 124194 / Chromium 125350 toward 24 November.
- NoisyTCP CVE-2026-99288 only if CISA or a major vendor moves it from remote DoS to exploited.
- BO Team / Log4Shell only if a current supply-chain payload is confirmed, not a 2025 IR recap.
- Cursor svc-telemetry only if Cursor revises the 5,000 ceiling or the 1 November containment date.
Sources
- FBI / IC3 — FLASH-20261113-001, Threat Overview of ARTEX Autonomous Penetration Testing Framework, 13 Nov 2026
- FBI / IC3 — Industry Alerts listing FLASH-20261113-001
- Amazon Threat Intelligence / CJ Moses — Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents, 13 Nov 2026
- Cloudflare — Adapting to AI penetration testing orchestrators, 13 Nov 2026
- BleepingComputer — FBI warns ARTEX AI hacking tool is powering global cyberattacks, 15 Nov 2026
- The Hacker News — Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys, 16 Nov 2026
- CISA — KEV JSON feed catalog 2026.11.13
- CISA — Adds Two Known Exploited Vulnerabilities to Catalog, 13 Nov 2026
- RWP Ventures — Cursor Fireworks marketplace Daily Top 2026-11-15
- RWP Ventures — Check Point DynamicCLI KEV Daily Top 2026-11-14
- RWP Ventures — Storm-3358 ARTEX financial Daily Top 2026-11-08
- RWP Ventures — LiteLLM cook.pth Daily Top 2026-10-25
- RWP Ventures — ConnectWise SkyKick OAuth Daily Top 2026-10-17