HighCVE-2025-3248CVE-2026-28797 5 Hunt LeadsIOCsRemote AttackerThe Hacker News

AI Agents Are Raiding Exposed AI Apps for Keys, Amazon’s Decoys Show

Amazon's MadPot decoys captured ARTEX agents' own records of hunting exposed LiteLLM, Dify, One API and other AI apps for provider keys, and of deepfake attempts against crypto exchange ID checks.

Amazon set out decoys dressed up as AI model services, and attackers’ AI agents handed over their own campaign notes. The Hacker News reports that Amazon’s threat intelligence team captured records showing agents, most likely running on the open-source ARTEX pentest framework, hunting exposed AI apps such as LiteLLM, One API and Dify for the provider keys stored inside, then using the stolen keys to pay for the AI that powered the hunt. Amazon says it can’t verify the outcomes the agents reported.

For defenders

What Amazon saw: from Oct. 3 to Nov. 5, 2026, sensors in Amazon’s MadPot decoy network posed as LLM chat-completion APIs. ARTEX agents sent them requests that carried whole campaign histories: operator instructions, tool output and reported findings, per Amazon’s write-up. Amazon assesses with moderate confidence that the operators were separate China-nexus groups (Simplified Chinese prompts, China time zones, a rule to avoid mainland Chinese and government targets). No group is named. AWS customers who may be affected were notified.

The playbook: targets were LLM gateways (One API, New API, Veloera, Sub2API, LiteLLM) and AI app builders (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama). Entry points: unfinished installs anyone could claim, management endpoints with no login, default admin credentials such as LiteLLM’s default master key (T1078.001 (Valid Accounts: Default Accounts)), an isCliRequest header bypass in One API-family apps, CVE-2025-3248 in Langflow (fixed in 1.3.0, on KEV) and CVE-2026-28797 in RAGFlow (sandboxed in 0.25.0): T1190 (Exploit Public-Facing Application). They also used SSRF through URL-fetch features to reach cloud metadata (T1552.005 (Unsecured Credentials: Cloud Instance Metadata API)) and scraped secrets from JavaScript, source maps, debug routes and .env files (T1552.001 (Unsecured Credentials: Credentials In Files)). Stolen keys went into a shared pool to run models on someone else’s bill: T1496.004 (Resource Hijacking: Cloud Service Hijacking). A second campaign tried deepfake selfies and forged ID photos against two crypto exchanges’ identity checks; one vendor rejected every deepfake video attempt.

Cloudflare’s case: in a separate ARTEX incident, an attacker used a stolen, never-expiring account-level API token and created Worker scripts named artex-*, artex-pgrecon-*, artex-pgw-* and artex-do-*, with about 400 database dump requests in an hour, per Cloudflare. Talos publishes ARTEX Snort rules (Snort 3 7:76 to 7:78, Snort 2 1:67494 to 1:67496) and indicator files. Amazon’s IPs are below; four more it lists sit in Cloudflare’s shared 104.28.0.0/16 range and are left out of the block list.

Network (Amazon, Oct. 14 to Nov. 2, 2026)
31.22.91[.]156
150.136.38[.]110
128.242.115[.]131
216.176.239[.]231
212.107.28[.]51
103.142.140[.]158
125.103.213[.]41

Background on ARTEX: AI Agents Turn Secrets Leaked on GitHub Into Cloud Break-Ins in Minutes. Today’s SANS scans of LiteLLM are a related but separate story.

What it means for you

If you don’t run AI tools on your own servers, nothing changes for you today. If your business set up a self-hosted AI chat app or a gateway that holds keys for OpenAI, Anthropic or other providers, assume automated attackers are already looking for it. They don’t need anything clever: a default password or a forgotten setup page is enough, and the first sign may be a shocking AI bill.

Do this now

  1. Find every self-hosted AI app or model gateway you run and take it off the open internet, or put a login in front of it.
  2. Change default admin credentials (including LiteLLM’s master key) and finish any half-done installs.
  3. Rotate the provider API keys those apps hold, and set spending limits and alerts with each AI provider.
  4. Patch Langflow to 1.3.0 or later and RAGFlow to 0.25.0 or later.

Source article: Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys
Publisher: The Hacker News
Published: Nov. 16, 2026
URL: https://thehackernews.com/2026/11/amazon-uncovers-ai-agent-campaign.html

Executive Summary

Between Oct. 3 and Nov. 5, 2026, Amazon Threat Intelligence’s MadPot decoy sensors, configured to look like LLM inference APIs, received requests from AI agents that carried the agents’ own campaign records. Amazon assesses with moderate confidence that the agents ran on ARTEX, an open-source agentic penetration testing framework whose repository was pulled on Oct. 8 but survives in forks and a commercial ARTEX PRO, and that the operators were disparate China-nexus threat groups; no group name or vendor alias is assigned. The main campaign was industrial-scale theft of AI provider credentials: agents fingerprinted internet-exposed LLM gateways (One API, New API, Veloera, Sub2API, LiteLLM) and AI app builders (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama), worked a codified playbook of claimable uninitialized installs, unauthenticated management and provider-configuration endpoints, default administrator credentials such as the LiteLLM master key, an isCliRequest header bypass in One API-family systems, CVE-2025-3248 and CVE-2026-28797, SSRF to cloud metadata, and secrets in client-side JavaScript, source maps, debug routes and environment files, then validated each key and pooled it to fund further inference. Cloud credentials were used to enumerate storage, managed AI model services, secret stores and application identities. A second campaign attempted deepfake-based KYC bypass at two cryptocurrency exchanges. Separately, Cloudflare reported an ARTEX intrusion using a stolen account-level API token that created artex-named Worker scripts. Amazon could not verify reported outcomes. Hunt leads, false positive expectations, published rule references and indicators are detailed below.

ATT&CK chain: Reconnaissance → Initial Access → Credential Access → Discovery → Impact

Hunt Lead Grid

Hunt Lead Level Hunt Lead Name ATT&CK Risk of False Positives
1 (Junior) MadPot Campaign IP Sweep T1595.002 (Active Scanning: Vulnerability Scanning) Low. The lead anchors on seven specific addresses Amazon published with first and last seen dates; some sit on cloud or hosting ranges that may be reassigned, so hits outside Oct. 14 to Nov. 2, 2026 need review rather than automatic escalation.
1 (Junior) ARTEX-Named Cloudflare Worker Scripts T1078.004 (Valid Accounts: Cloud Accounts) Low. The lead anchors on Worker script names beginning with artex-, which no legitimate deployment is expected to use unless a team is running an authorized ARTEX test.
2 (Pro) AI Gateway Admin Access With Default or No Credentials T1078.001 (Valid Accounts: Default Accounts); T1190 (Exploit Public-Facing Application) Medium. Administrators use the same management endpoints; the lead anchors on access from outside administrative networks, on the default master key, on the isCliRequest header from remote clients and on setup routes after installation.
2 (Pro) AI App SSRF to Cloud Metadata T1552.005 (Unsecured Credentials: Cloud Instance Metadata API) Medium. AI apps fetch URLs by design; the lead anchors on fetches by the AI app process to the link-local metadata address or metadata hostnames, which user-supplied content should never need.
3 (Elite) Provider Key Validation and Reuse From New Origins T1496.004 (Resource Hijacking: Cloud Service Hijacking) Medium. Key usage legitimately grows and moves; the lead anchors on a provider key used from an origin outside the gateway’s own egress addresses, preceded by low-cost validation calls such as model listing.

MadPot Campaign IP Sweep

Amazon published the source addresses that delivered the agents’ requests to its MadPot sensors: five for the credential theft campaign (first seen Oct. 14, last seen Nov. 2, 2026) and six for the deepfake KYC campaign (Oct. 14 to 19). Seven are listed below; four others fall in Cloudflare’s 104.28.0.0/16 range, which is shared egress used by many legitimate clients, and are deliberately left out of the indicator list. The detection logic is a retrospective sweep of inbound connections to any AI application, model gateway or inference endpoint, and of outbound connections from them, for the seven addresses, focusing on requests to management, provider configuration, setup and model-listing routes.

Sweep web server, reverse proxy, load balancer and WAF logs, plus firewall and NetFlow records, back to at least Oct. 1, 2026, because the campaign ran from Oct. 3. Scope it to hosts running the products named in the playbook first, then widen to all internet-facing web services. Blocking these addresses now is reasonable but not a fix, since the agents rotate infrastructure and Talos links more than 2,500 addresses to ARTEX activity. Any hit against an AI gateway with a 200 response on a management or provider endpoint should be escalated as a likely key theft and handled by rotating every provider key the gateway holds.

31.22.91[.]156     # Campaign 1, Oct. 14 to 19
150.136.38[.]110   # Campaign 1, Oct. 31
128.242.115[.]131  # Campaign 1, Oct. 31
216.176.239[.]231  # Campaign 1, Nov. 2
212.107.28[.]51    # Campaign 2, Oct. 14
103.142.140[.]158  # Campaign 2, Oct. 17
125.103.213[.]41   # Campaign 2, Oct. 18 to 19

Hunt Lead Level: 1 (Junior Analyst)

ATT&CK: Reconnaissance → T1595.002 (Active Scanning: Vulnerability Scanning)

Sources: “Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents”, AWS Security Blog (CJ Moses), Nov. 13, 2026, https://aws.amazon.com/blogs/security/amazon-threat-intelligence-observes-threat-actors-running-cybercrime-campaigns-with-artex-ai-agents/ (additional source, same campaign); “Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys”, The Hacker News, Nov. 16, 2026 (source article)

ARTEX-Named Cloudflare Worker Scripts

In Cloudflare’s ARTEX incident, the attacker held a compromised account-level API token that had been used for CI/CD, had no IP restriction and never expired, most likely stolen through a third-party cloud provider. With it, new Worker scripts appeared named artex-*, artex-pgrecon-*, artex-pgw-* and artex-do-*, each name matching a task, and activity ran at more than 69 requests per minute across several parts of the account, including roughly 400 database dump requests in one hour and attempts to establish persistence and validate session tokens for other third-party apps. The detection logic is to query Cloudflare audit logs for Worker script creation or update events whose script name starts with artex-, and for any Worker created by an API token rather than a dashboard user outside a normal deployment pipeline.

Use Cloudflare account audit logs (via Logpush or the API) and your CI/CD system’s deployment records to separate pipeline deploys from ad hoc ones. Sweep back 90 days, since Cloudflare says it investigated multiple ARTEX-related cases over eight weeks. Do not rely on the artex- prefix alone in the long run, because names are trivial to change; pair it with token-level anomalies such as a token used from a new IP or at an unusual request rate. Any artex-named Worker you did not create is a confirmed compromise: revoke the token, delete the scripts, review database and storage access in the window, and reissue tokens with an expiry and client IP restrictions.

Hunt Lead Level: 1 (Junior Analyst)

ATT&CK: Initial Access, Persistence → T1078.004 (Valid Accounts: Cloud Accounts)

Sources: “Adapting to AI penetration testing orchestrators”, Cloudflare Blog, Nov. 2026, https://blog.cloudflare.com/adapting-to-ai-penetration-testing-orchestrators/ (additional source, same ARTEX tooling, as relayed by the source article); source article (above)

AI Gateway Admin Access With Default or No Credentials

The agents’ playbook mapped each fingerprinted product to its known entry point: uninitialized installs that could be claimed by whoever completes setup first, anonymous management and provider-configuration endpoints, default administrator credentials such as LiteLLM’s default master key, an authentication bypass via the isCliRequest header in One API-family systems (One API, New API, Veloera, Sub2API), and known flaws CVE-2025-3248 (Langflow, missing authentication leading to code execution) and CVE-2026-28797 (RAGFlow, template injection by any registered user, with self-registration on by default). The goal on every product was to read the upstream provider keys the app aggregates. The detection logic is to alert when a management, channel, provider or key-listing endpoint on these products returns success to a client outside administrative networks; when a request authenticates with the product’s default master key; when any remote request carries an isCliRequest header; and when setup or initialization routes are hit after the install was completed.

Use reverse proxy or WAF logs in front of the apps (most of these products log little themselves), plus the application’s own audit or admin logs where available. Scope to every self-hosted AI gateway and app builder, including shadow deployments on developer and data science hosts, which asset discovery for port 4000 (LiteLLM’s default), 3000 and 7860 often finds. Changing the default key and putting the app behind authentication or a VPN are the controls; hash- or IP-blocking is the wrong primary control here, because the agents rotate infrastructure and the weakness is configuration. A successful hit means the provider keys are burned: rotate every key the app holds and check each provider’s usage for the period, then use the final lead to look for reuse.

Hunt Lead Level: 2 (Pro)

ATT&CK: Initial Access → T1078.001 (Valid Accounts: Default Accounts); Initial Access → T1190 (Exploit Public-Facing Application)

Sources: AWS Security Blog (above), additional source; source article (above)

AI App SSRF to Cloud Metadata

The playbook included server-side request forgery through the URL-fetch features common in AI app builders (document loaders, web retrieval tools, image fetchers) to reach cloud instance metadata services and pull the host’s cloud credentials. When agents obtained cloud credentials, Amazon says they validated identities through each provider’s token and identity services, enumerated storage, assessed permission scope and went after managed AI model services, secret and parameter stores, and application identities, prioritizing premium model access. The detection logic is to flag any outbound request from an AI application process or container to the link-local metadata address (169.254.169.254) or a provider metadata hostname, and any subsequent use of that instance’s role credentials from an address that is not the instance itself.

Use EDR or container runtime network telemetry, VPC flow logs, and cloud audit logs such as AWS CloudTrail for role credential use, joined by instance identity. Enforce IMDSv2 with a hop limit of 1 on instances running these apps, which defeats most simple SSRF, and restrict the app’s egress to the providers it needs. Allowlisting the metadata address for the app process would be the wrong move, because the AI app itself has no reason to request metadata from user-driven fetches; the cloud SDK’s own credential refresh comes from a distinct process or library path you can baseline. Instance credentials used from an outside address are a confirmed theft and warrant immediate role session revocation.

Hunt Lead Level: 2 (Pro)

ATT&CK: Credential Access → T1552.005 (Unsecured Credentials: Cloud Instance Metadata API)

Sources: AWS Security Blog (above), additional source; source article (above)

Provider Key Validation and Reuse From New Origins

Every stolen key was validated and graded (L1 lead, L2 confirmed working call, L3 quantified impact such as readable data or spendable balance) before going into a managed pool that fed ARTEX’s own model gateway, so the attackers’ inference ran on victims’ bills. Keys exposed publicly for more than 30 days were treated as low value, which means fresh leaks are used fastest. The detection logic correlates two sources: your gateway’s own egress addresses toward each AI provider, and the provider’s usage or audit records for the same keys. Alert when a key is used from an origin outside that set, especially when the first calls are cheap validation requests (listing models, minimal completions) followed by a sustained burst of premium-model usage.

Use AI provider usage dashboards, organization audit logs and billing exports where the provider offers per-key and per-IP detail, combined with firewall or proxy logs of gateway egress, and baseline normal per-key spend for two to four weeks. Spend alerts and hard limits per key are the fastest practical control and catch theft even where per-IP data is missing. Simply rotating a key after a bill spike, without finding how it leaked, is the wrong response, since the same exposed app will leak the replacement. Treat any confirmed reuse as an incident on the gateway that held the key, and the pattern generalizes to any stolen SaaS API key, not only LLM providers.

Hunt Lead Level: 3 (Elite)

ATT&CK: Impact → T1496.004 (Resource Hijacking: Cloud Service Hijacking)

Sources: AWS Security Blog (above), additional source; source article (above)

Published rules: the source article cites Cisco Talos Snort 3 rules 7:76, 7:77 and 7:78 and Snort 2 rules 1:67494, 1:67495 and 1:67496 for ARTEX (7:78 and 1:67496 are off by default and detect traces of the norma agent harness). The rule text is distributed through the Snort rulesets and is not reproduced in the article, so it is not included here; see Talos’ “ARTEX and beyond” post.

Hash Appendix

No file hashes are published in the source article, Amazon’s post or Cloudflare’s post for this activity; the indicators are network addresses and Worker script names. OTX AlienVault and VirusTotal were not queried because there is no file to pivot from.

References Appendix

  • “Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys”, The Hacker News, Nov. 16, 2026, https://thehackernews.com/2026/11/amazon-uncovers-ai-agent-campaign.html (source article)
  • “Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents”, AWS Security Blog (CJ Moses), Nov. 13, 2026, https://aws.amazon.com/blogs/security/amazon-threat-intelligence-observes-threat-actors-running-cybercrime-campaigns-with-artex-ai-agents/
  • “Adapting to AI penetration testing orchestrators”, Cloudflare Blog, Nov. 2026, https://blog.cloudflare.com/adapting-to-ai-penetration-testing-orchestrators/
  • “ARTEX and beyond”, Cisco Talos, https://blog.talosintelligence.com/artex-and-beyond/ (rule SIDs as cited by the source article)
  • Cisco Talos IOC repository, https://github.com/Cisco-Talos/IOCs/tree/main/2026/11
  • MITRE ATT&CK Enterprise technique pages linked above, MITRE, https://attack.mitre.org/

CVE at a glance

CVE-2025-3248CriticalKEVPublic exploitLangflow
CVSS
9.8 Critical (v3.1)
EPSS
100% chance of exploitation in the next 30 days (higher than 99.9% of CVEs)
Exploited in the wild
Yes, on CISA’s KEV list since May 5, 2025; used by ransomware
Public exploit
Yes, reported in a Metasploit module, Exploit-DB, a Nuclei template and 29 GitHub repositories

Full CVE details

CVE-2026-28797HighPublic exploitinfiniflow ragflow
CVSS
8.7 High (v4.0)
EPSS
0.6% chance of exploitation in the next 30 days (higher than 48.4% of CVEs)
Exploited in the wild
Not on CISA’s KEV list
Public exploit
Yes, reported in 1 GitHub repository

Full CVE details

Data from NVD, FIRST EPSS and CISA KEV, refreshed automatically. We report whether public exploit code exists but never link to it: fake proof-of-concepts that infect researchers are common.

More from the Roundup