TAU-HOME.COM
LOADING

Hacker Agents Catch AI Model Credential Theft: Amazon MadPot and Google GTIG Analyses

Amazon Threat Intelligence's inference API honeypot MadPot has captured ARTEX-based Chinese hacker agents attempting AI model credential theft, while Google GTI

tau · November 22, 2026

#Security #AIAgents #ThreatIntelligence #ZeroDay

Hacker Agents Catch AI Model Credential Theft: Amazon MadPot and Google GTIG Analyses

Amazon Threat Intelligence announced that its large-scale inference API honeypot, "MadPot," captured credential theft attempts targeting AI models by Chinese hacker agents operating on the open-source penetration framework "ARTEX." Concurrently, the Google Threat Intelligence Group (GTIG) exposed "exploit_agent," an autonomous framework that automatically compares code before and after software security patches to generate and execute zero-day exploits, sending shockwaves through the security community.

Threat intelligence report screen covering AI security threats and hacker agent analysis

Image source: https://agentsecurity.com/news/cyber-attack-news-2026-11-20

This incident marks a clear turning point where artificial intelligence agents are being actively harnessed not just by defenders, but as formidable attack tools by adversaries. According to threat intelligence reports published on November 20, 2026, attackers have organized automated agent swarms to scour internet-exposed AI applications and Large Language Model (LLM) gateways worldwide, using stolen API keys to fund subsequent rounds of credential hunting in a self-sustaining loop.

ARTEX-Based Hacker Agent Reconnaissance and Theft Captured by MadPot

Analyzing data collected by its MadPot sensor network between October 3 and November 5, 2026, Amazon Threat Intelligence observed AI agents hunting for upstream AI provider authentication credentials. These sensors impersonated popular LLM inference APIs, prompting visiting agents to leave behind comprehensive campaign records including system prompts, tool execution histories, shared memory, and reported findings.

The investigation revealed that these agents were built upon ARTEX, an open-source agentic penetration testing framework, and were operated by distinct China-nexus threat clusters. Their attack chain followed a highly structured progression:

  • Mission Parameters and Guardrails: Operators provided agents with standing objectives, targeting rules, and boundaries—such as excluding mainland China targets and government domains—while routing all outbound traffic through proxies.
  • Exploration Targets: Agents scanned internet-exposed AI applications that aggregate upstream provider keys, focusing on LLM gateways and proxies (such as One API, New API, and LiteLLM) alongside AI app builders (including Dify, Langflow, RAGFlow, and Open WebUI).
  • Multi-Vector Discovery: They leveraged internet scanning services, certificate transparency logs, GitHub repositories, package registries (npm and PyPI), container image layers, and Telegram channels to unearth leaked secrets.
  • Grading and Triage: Validated keys were categorized into leads (L1), verified working API calls (L2), and quantified impact metrics (L3) for managed inventory pools.

Amazon noted, however, that it did not observe the activity directly on victim systems as it transpired and could not independently verify the outcomes reported by the agents.

Google GTIG Exposes Autonomous Patch-to-Exploit Framework and Additional Threats

In a simultaneous disclosure, the Google Threat Intelligence Group (GTIG) reported that it had identified threat actors using AI frameworks to fully automate vulnerability discovery and exploitation without human intervention, centered around a framework named exploit_agent.

Discovered on an exposed staging server, this multi-agent framework connects to multiple frontier and local models to run continuous differential analysis on popular web application software. It compares source code before and after patch announcements, identifies potential vulnerabilities, drafts exploits, and executes them automatically. The compromised server also housed stolen victim data and post-compromise reconnaissance scripts.

In addition, GTIG detailed campaigns where a North Korean-linked cluster (such as MIDNIGHT NEPTUNE) targeted cryptocurrency organizations using fake recruiter profiles. After initial compromise, the threat actor modified .claude/settings.json files within internal code repositories so that developers launching Claude Code would automatically download the SOMBERMEME backdoor. Furthermore, malicious versions of the agent-framework package (1.26.1 and 1.26.2) on PyPI were found to contain embedded credential stealers before being swiftly removed.

Implications for Production Environments and AI Infrastructure Security

These threat intelligence disclosures serve as a sobering reminder that while AI agents are potent tools for elevating developer productivity, they can become dangerous force multipliers in the hands of malicious operators capable of bypassing security controls at unprecedented speed and scale.

With the Cybersecurity and Infrastructure Security Agency (CISA) adding the LiteLLM vulnerability (CVE-2026-40217) to its Known Exploited Vulnerabilities (KEV) catalog and highlighting its connection to LLMJacking intrusions, organizations running AI agents in production must take immediate hardening steps:

  1. Verify Agent Configuration Files: Thoroughly inspect cloned external repositories and local configuration files—such as .claude/settings.json—for hidden startup hooks or unauthorized remote download commands.
  2. Isolate Exposed LLM Gateways: Restrict access to administrative endpoints and API gateways (such as One API, LiteLLM, and RAGFlow) that lack proper authentication, and immediately change default master keys.
  3. Rotate and Monitor API Keys: Implement strict monitoring for anomalous consumption patterns across cloud and AI model provider tokens, and routinely rotate keys to mitigate potential exposure risks.

Sources

Impact-Site-Verification: 9b3ebd10-6c45-44cf-a014-5b8c19b7b285