← Voltar ao feed

China-linked actors use ARTEX AI pentest agents to steal LLM API keys, powering the agents with the stolen keys, per Amazon

4 fonte(s) · 4 atualização(ões) · publicado 14/11/2026 04:19 · coletado há 23 horas

Matéria consolidada. Reúne 4 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que as 4 fontes trouxeram em 17/11/2026 00:01; os links antigos trazem para cá.

Fontes: theregister.com ↗ securityaffairs.com ↗ thehackernews.com ↗ gbhackers.com ↗

Amazon Threat Intelligence reported that several China-linked groups (moderate confidence) used the open-source agentic penetration-testing framework ARTEX to scan exposed AI applications and steal AI provider keys (securityaffairs.com, thehackernews.com). Amazon's MadPot decoy sensors, configured to look like LLM inference APIs, received requests carrying the agents' own campaign records between 3 October and 5 November 2026 (securityaffairs.com, theregister.com). The FBI issued a Flash alert on ARTEX abuse, and the abuse reportedly targets banks, governments, energy and healthcare providers. Some intrusions go from initial targeting to privileged access in under three hours (theregister.com). The stolen keys feed the framework's model gateway and power the agents hunting for more keys, which creates a self-sustaining inference loop (gbhackers.com, securityaffairs.com). The credential theft covered AI model API keys and cloud credentials at "industrial scale" (theregister.com). Targets include LiteLLM, One API, Dify, Langflow and RAGFlow, and some also include cloud services. A second, related campaign used real-time deepfakes and forged ID photos against KYC liveness checks at two cryptocurrency exchanges. Amazon did not observe the attacks live and cannot verify the outcomes the agents reported. It notified AWS customers who may have been affected (thehackernews.com).

Entenda a falha

A vulnerabilidade

Two CVEs are listed. CVE-2025-3248 is a remote code execution flaw in Langflow (per gbhackers). CVE-2026-28797 affects RAGFlow, and the articles give no type or impact for it. No article gives affected or fixed versions for either, and the status list of playbook flaws in thehackernews.com is missing from the text. Both are entries in the ARTEX playbook, not new techniques. The campaign mostly abused weaknesses with no CVE: LiteLLM's default master key, unfinished installs, unauthenticated admin and provider-settings endpoints, an authentication bypass via the isCliRequest header in One API-family systems, SSRF through URL-fetch features, and secrets in client-side files. The attacker needs network reach to an exposed instance and gets the model-provider keys stored there.

Como funciona

ARTEX agents (planner, worker, verification, compression) share a persistent knowledge graph. An operator sets a goal, and the planner splits it into tasks for workers that run security tools and ask an LLM for each next step. The agents find exposed AI gateways and platforms through internet scanners, certificate transparency logs, GitHub, npm, PyPI, container layers, paste sites and Telegram. For each product they follow a playbook: default keys, open admin pages, the isCliRequest bypass, the two CVEs, SSRF to cloud metadata, and secrets in JavaScript, source maps, debug routes and env files. Keys are verified and graded L1, L2 or L3, then added to a shared pool and fed to the framework's model gateway, which pays for more agent inference.

Por que é vulnerável

Amazon blames basic hygiene failures: default or hardcoded credentials, keys left in code and public artifacts, unauthenticated interfaces and inference endpoints, and unpatched software. These platforms concentrate keys from many providers, so one weak admin page exposes many accounts. Keys leaked in places outside automated revocation, such as GitHub event archives, stay usable. In the Cloudflare cases, the root cause was static CI/CD tokens that never expire, with no IP restriction. The articles do not explain the root cause of CVE-2025-3248, CVE-2026-28797 or the isCliRequest bypass, nor what the fixes change.

Como foi descoberta

Amazon Threat Intelligence found the campaign through MadPot decoy sensors posing as LLM inference APIs. From 3 October to 5 November, ARTEX agents sent them requests carrying their own campaign records. Amazon did not see the attacks live and cannot verify the outcomes. It assesses with moderate confidence that several China-linked groups used ARTEX, notified affected AWS customers and published IPs seen between 14 October and 2 November. Separately, Cloudflare incident responders saw ARTEX use a compromised CI/CD token, and an FBI Flash alert summarized 11 earlier reports of use since August. The ARTEX developer pulled its GitHub repository on 8 October, but forks and a commercial version remain.

PoC existente, privada

O exploit existe, mas quem o tem não o publicou.

The articles show exploitation in the wild: the ARTEX playbook includes checks for the Langflow and RAGFlow CVEs and was used against exposed platforms. ARTEX itself is open source, with forks and archived copies still circulating, and Censys counted 1,465 exposed ARTEX consoles on 4 November. No article says a standalone exploit for either CVE was published, so access is treated as private. Exposed instances should be treated as at real risk: patch, rotate keys, remove default credentials and restrict endpoints now.

Explicação gerada por IA a partir da matéria e dos artigos, em 17/11/2026 00:17.

2 exploits públicos

Há código de exploração publicado no Exploit-DB para CVE-2025-3248. O código não é reproduzido aqui: o link leva à fonte, e a explicação foi escrita pela IA a partir do arquivo publicado.

  1. exploit-db.com · EDB-52364 ↗ · aplicação web · multiple · pub. 16/07/2025 não verificado

    Langflow 1.2.x - Remote Code Execution (RCE)

    This script targets Langflow 1.2.x and earlier through its code-validation API endpoint, which evaluates submitted Python code on the server. An unauthenticated attacker with network access to the Langflow web service can make it run arbitrary operating system commands with the privileges of the Langflow process. The script sends one crafted request and prints the command output returned in the error field of the response. It matches CVE-2025-3248 and looks like a working, simple exploit, though Exploit-DB has not verified it. Upgrading Langflow and not exposing it to untrusted networks mitigates it.

    CVE-2025-3248 · autor: Raghad Abdallah Al-syouf · explicação gerada por IA
  2. exploit-db.com · EDB-52262 ↗ · remoto · multiple · pub. 18/04/2025 não verificado

    Langflow 1.3.0 - Remote Code Execution (RCE)

    This is a small Python script for CVE-2025-3248 in Langflow, the AI workflow builder. It abuses the web API's code validation feature, which reaches the Python interpreter without requiring login, so an attacker only needs network access to a Langflow server running a version before 1.3.0. The script sends a crafted function definition that makes the server run an operating-system command chosen by the attacker, then reads the output back from the error message. It looks like a working, simple exploit that matches the CVE, though Exploit-DB has not verified it.

    CVE-2025-3248 · autor: VeryLazyTech · explicação gerada por IA

Detalhes técnicos

  • Architecture: planner, worker, verification and compression agents share a persistent knowledge graph (securityaffairs.com).
  • Targets: API gateways and proxies (One API, New API, Veloera, Sub2API, LiteLLM) and AI platforms (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama) (securityaffairs.com, theregister.com).
  • Initial access: default admin credentials such as the LiteLLM default master key, unfinished installs that can be taken over, and unauthenticated management endpoints. Known flaws include CVE-2025-3248 (Langflow) and CVE-2026-28797 (RAGFlow). Other routes are an authentication bypass via the isCliRequest header in One API-family systems, SSRF through URL-fetch features to reach cloud metadata services, and secrets in client-side JavaScript, source maps, debug routes and env files (securityaffairs.com, thehackernews.com). The agents succeeded where default or hardcoded credentials, keys in code, unauthenticated interfaces and unpatched software were present (theregister.com).
  • Key discovery: internet scanners, certificate transparency logs, GitHub, npm, PyPI, container layers, paste sites and Telegram. Credentials were graded L1 (lead), L2 (working call) and L3 (measured impact) (securityaffairs.com).
  • Cloudflare observations: ARTEX used an account-level CI/CD API token with no IP restriction or expiry. It created Worker scripts named artex-, artex-pgrecon-, artex-pgw-* and artex-do-*, made more than 69 requests per minute and 400 database dump requests in an hour, and attempted persistence. The root cause was static CI/CD tokens that never expire (thehackernews.com).
  • KYC campaign: the agents reverse-engineered a vendor liveness check. They built a live deepfake pipeline at 11.5 fps on a single consumer GPU, with face placement at 7 pixels, and forged EXIF/camera metadata and JPEG compression tables on ID photos. The vendor's server rejected the deepfake every time it was recorded (securityaffairs.com, theregister.com).
  • Attribution basis: Simplified Chinese prompts, China-linked time zones, targets matching Chinese interests, and instructions to avoid mainland China and government domains (securityaffairs.com).
  • Government and vendor context: the FBI Flash alert reportedly summarizes 11 earlier reports of ARTEX use since August (thehackernews.com). The FBI cited August attacks on South Korean financial institutions (about 67,000 people affected) and a September supply-chain attack on ConnectWise Cloud Backup customers that compromised more than 2,250 Microsoft tenants (theregister.com). Cisco Talos published Snort rules 7:76-78 and 1:67494-67496 (thehackernews.com).
  • Tool availability: the original developer removed the GitHub repository in October, but forks and a commercial version remain. Censys counted 1,465 exposed ARTEX consoles on 4 November (theregister.com).

Remediação

  • Treat AI applications, model gateways and inference endpoints as production systems.
  • Remove default and hardcoded credentials, finish new installs, and patch known flaws such as CVE-2025-3248 and CVE-2026-28797.
  • Scan for embedded secrets and rotate any exposed LLM API keys.
  • Give Cloudflare and other CI/CD tokens an expiry and IP restrictions, and restrict management endpoints.
  • Restrict sensitive services to trusted networks, private connectivity or dedicated admin paths, and put enterprise applications behind zero-trust access (theregister.com).
  • Close unauthenticated inference endpoints.
  • Hunt for artex-* Workers and for the IP addresses Amazon published, which were seen between 14 October and 2 November.
  • Monitor for unexpected spend or unusual usage, and deploy the Cisco Talos Snort rules.

Relatos divergentes

  • Attribution confidence: gbhackers.com describes the actors flatly as Chinese-nexus. thehackernews.com and securityaffairs.com, relaying Amazon, say it is assessed with only moderate confidence and involves several separate China-linked groups. The Amazon-based framing is more authoritative.
  • Timeframe: gbhackers.com gives October to November 2026 generally. securityaffairs.com and theregister.com give the precise MadPot collection window of 3 October to 5 November, and the FBI alert covers use since August (thehackernews.com).
  • Outcomes: gbhackers.com states the deepfake injections were not successful. securityaffairs.com is more careful: Amazon did not observe the attacks live and cannot confirm the agents' reported results, though the vendor's server rejected the deepfake each time it was recorded. theregister.com also reports the server-side rejection in every attempt.
  • FBI alert: thehackernews.com says the FBI "reportedly" issued the alert. theregister.com, the more recent source, reports the Flash alert directly alongside Amazon's findings.

Linha do tempo

  1. theregister.com ↗ · pub. 16/11/2026 16:52 · coletado 16/11 17:00

    ARTEX open-source AI pentest framework abused to automate attacks on critical sectors and steal AI credentials, FBI and AWS warn

    ler a análise desta fonte

    The FBI (Flash alert) and Amazon Threat Intelligence warn that attackers are abusing the open-source AI penetration testing framework ARTEX to automate intrusions against banks, governments, energy and healthcare providers. Between October 3 and November 5, Amazon's MadPot honeypots detected two campaigns. One stole AI model API keys and cloud credentials at "industrial scale". The other built real-time deepfake video and forged ID photos to defeat KYC checks at two cryptocurrency exchanges. The vendor's server-side verdict rejected the deepfake in every attempt. The FBI says some intrusions go from initial targeting to privileged access in under three hours. This is a report from the article alone, because no web research was performed.

    Technical Details

    • Credential campaign: AI agents stole API keys and fed them into ARTEX's model gateway to power more agents. Targets included LLM gateways and proxies (One API, New API, Veloera, Sub2API, LiteLLM) and AI app builders (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama).
    • KYC campaign: agents reverse-engineered a vendor's liveness protocol. They built a deepfake pipeline on a single consumer GPU running at 11.5 fps, and forged camera metadata and JPEG compression tables.
    • The agents succeeded where default or hardcoded credentials, keys in code, unauthenticated interfaces and unpatched software were present.
    • The FBI cited August attacks on South Korean financial institutions (about 67,000 people affected) and a September supply-chain attack on ConnectWise Cloud Backup customers that compromised more than 2,250 Microsoft tenants.
    • The original developer removed the GitHub repository in October. Forks and a commercial version remain, and Censys counted 1,465 exposed ARTEX consoles on November 4.

    Remediation

    Restrict sensitive services to trusted networks, private connectivity or dedicated admin paths. Put enterprise applications behind zero-trust access. Fix basic hygiene: default credentials, exposed keys, unauthenticated inference endpoints and unpatched software.

  2. securityaffairs.com ↗ · pub. 16/11/2026 06:55 · coletado 16/11 07:00

    China-linked actors use ARTEX AI agents to steal LLM API keys and test deepfake KYC bypasses at crypto exchanges, per Amazon

    This article adds MadPot collection details (3 October to 5 November), the agent architecture, the credential grading scheme, the isCliRequest bypass, KYC deepfake specifics (11.5 fps pipeline, 7-pixel face placement, forged EXIF) and Amazon's caveat that it did not observe the attacks live.

    ler a análise desta fonte

    This article adds MadPot collection details (3 October to 5 November), the agent architecture, the credential grading scheme, the isCliRequest bypass, KYC deepfake specifics (11.5 fps pipeline, 7-pixel face placement, forged EXIF) and Amazon's caveat that it did not observe the attacks live.

    Amazon Threat Intelligence reported that criminals, assessed with moderate confidence to be several China-linked groups, ran offensive campaigns with the open-source agentic penetration-testing framework ARTEX. Between 3 October and 5 November, Amazon's MadPot decoy sensors, configured to look like LLM inference APIs, received requests that carried the agents' own campaign records. The largest campaign stole AI credentials, and the stolen keys fed the framework's model gateway, which powered the agents hunting for more keys. A second campaign used real-time deepfakes and forged ID photos against KYC checks at two cryptocurrency exchanges. Amazon did not observe the attacks live and could not confirm that the agents met their goals.

    Technical Details

    • Planner, worker, verification and compression agents shared a persistent knowledge graph.
    • Targets were API gateways and proxies (One API, New API, Veloera, Sub2API, LiteLLM) and AI platforms (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama).
    • Techniques included default keys (such as LiteLLM's default master key), exposed admin pages, and known flaws CVE-2025-3248 (Langflow) and CVE-2026-28797 (RAGFlow). They also included an isCliRequest header authentication bypass in One API-related systems, SSRF against cloud metadata, and secrets exposed in JavaScript, source maps and env files.
    • Discovery used internet scanners, certificate transparency logs, GitHub, npm, PyPI, container layers, paste sites and Telegram. Credentials were graded L1 (lead), L2 (working call) and L3 (measured impact).
    • In the KYC campaign, the agents reverse-engineered a vendor liveness check. They built a live deepfake pipeline at 11.5 fps on a single consumer GPU and forged camera metadata on ID photos. The vendor's server rejected the deepfake every time it was recorded.
    • Attribution rests on Simplified Chinese prompts, China-linked time zones, targets matching Chinese interests, and instructions to avoid mainland China and government domains.

    Remediation

    Treat AI applications, model gateways and inference endpoints like production systems. Remove default and hardcoded credentials, rotate exposed keys, patch, and monitor unexpected costs and unusual usage. Amazon published the IP addresses seen between 14 October and 2 November for hunting.

  3. thehackernews.com ↗ · pub. 16/11/2026 04:34 · coletado 16/11 05:00

    Amazon: China-linked AI agents (ARTEX) hunt LiteLLM, Dify, One API for stored AI provider keys

    ler a análise desta fonte

    Amazon Threat Intelligence captured records from its MadPot decoy network showing attacker AI agents, assessed with moderate confidence to run on the open-source pentest tool ARTEX, scanning internet-exposed AI apps such as LiteLLM, One API, Dify, Langflow and RAGFlow. They steal and verify AI provider keys and pool them, and the stolen keys then power the agents' own models. Amazon assesses, with moderate confidence, that the operators are separate China-linked groups. It cannot verify the outcomes the agents reported, and it notified AWS customers who may have been affected. This summary draws on the article alone. No vendor advisory or other source was consulted.

    Technical Details

    • The playbook lists default admin credentials (such as the LiteLLM default master key), unfinished installs that an outsider could take over, and known flaws in Langflow and RAGFlow.
    • It also lists unauthenticated management endpoints, an authentication bypass through the isCliRequest header in One API family systems, SSRF through URL-fetch features to reach cloud metadata services, and secrets left in client-side JavaScript, source maps, debug routes and env files.
    • Agents searched for leaked keys on GitHub, npm, PyPI, container layers, paste sites, Telegram, certificate transparency logs and scanning services.
    • Cloudflare saw ARTEX use an account-level CI/CD API token with no IP restriction or expiry. It created Worker scripts named artex-, artex-pgrecon-, artex-pgw-* and artex-do-*, ran more than 69 requests per minute, made 400 database dump requests in an hour, and tried persistence. The root cause was static CI/CD tokens that never expire.
    • The FBI issued a FLASH alert summarizing 11 earlier reports of ARTEX use since August.
    • A second campaign used deepfakes against two cryptocurrency exchanges' identity checks.
    • Cisco Talos published Snort rules 7:76-78 and 1:67494-67496.

    Remediation

    • Change default credentials and finish new installs.
    • Scan for embedded secrets and rotate any that are exposed.
    • Give Cloudflare tokens an expiry and IP restrictions.
    • Hunt for the artex-* Workers and the published IPs.
    • Monitor for unexpected spend or unusual usage.
    • Deploy the Snort rules.
  4. gbhackers.com ↗ · pub. 14/11/2026 04:19 · coletado 14/11 06:00 primeiro registro

    Chinese-Nexus Hackers Use Stolen LLM API Keys to Fuel ARTEX AI Agents in Credential Theft Campaign

    ler a análise desta fonte

    Between October and November 2026, Chinese-nexus threat actors leveraged the ARTEX open-source agentic penetration-testing framework to steal large language model (LLM) API credentials. These stolen keys were used to power the ARTEX AI agents, creating a self-sustaining inference loop that enabled continued credential theft without requiring additional operator resources. Targets included multiple AI gateway platforms and cloud services, exploiting default keys, misconfigurations, and vulnerabilities such as CVE-2025-3248 and CVE-2026-28797. Operations involved extensive reconnaissance, exploitation, and validation steps, with specific geographical and operational indicators pointing to China. In parallel, a related campaign employed deepfake injections targeting cryptocurrency exchange KYC liveness checks, though without successful bypass. The campaign underscores the need for patching, credential rotation, and endpoint restriction to mitigate risks associated with exposed LLM API keys and management interfaces.