GTIG reports live intrusions using autonomous multi-agent frameworks like ARTEX and PentAGI, while LLMJacking and AI-assisted influence operations scale across crime, espionage, and IO. Adversaries automate differential patch analysis and n-day weaponization; stolen LLM credentials and cloud quotas are resold via LLM Mixers. PRC-nexus actors stage data on model file-hosting backends.
ai-agent-security
19 posts
GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold Tracing the Unrestricted AI Ecosystem: From Public Repositories to Hosted Services Abliterated open-weight models sold through hosted APIs lower attackers' skill barrier and expose defenders to unvetted inference and data risks. LevelBlue found 290 GitHub repositories referencing abliteration, nearly 9,000 Hugging Face checkpoints, and commercial gateways selling reduced-refusal models per token. The risk lies in unapproved inference services and coding agents with credentials and production access, not the models themselves.
ARTEX and beyond: AI-enabled penetration testing platforms China-nexus and criminal actors use ARTEX, an open-source AI penetration-testing framework released July 2026, to automate reconnaissance, credential theft, and intrusions. Talos identified over 2,500 ARTEX-associated C2 IPs hosting tools including VShell, AdaptixC2, Mimikatz, and Cobalt Strike. Campaigns attributed to VAPOR PANDA, Storm-0940, and Storm-3358 span espionage and financial theft; ARTEX ships no C2 but extends through MCP tool servers.
Inside the loop: Five AI attack frameworks that never ask permission Five agentic offensive AI frameworks — ARTEX, CyberStrikeAI, Cairn, PentAGI, HexStrike — lack code-enforced authorization and have appeared in intrusions, including Storm-0940's ARTEX use. The frameworks run scanners, shells, and cloud CLIs in two- to 25-second bursts, with approval prompts disabled by default. Only .gov and .edu blocklists restrict targets; ARTEX and CyberStrikeAI switch proxies, User-Agents, and forwarding headers when blocked.
Trained to betray: How post-training supply chain attacks hijack coding agents | Microsoft Security Blog Microsoft research shows poisoned fine-tuning data or attacker-controlled open-weight checkpoints can implant coding-agent backdoors, triggering unsafe tool calls at up to 94.8% success on unseen repositories. Poisoned examples teach models to discover and misuse available command-execution tools. Backdoors survive paraphrased triggers and different agent harnesses.
Attackers Use Autonomous AI to Target Semiconductor, Defense and Civil Society Sectors Chinese-nexus CL-STA-1238 used the open-source Cairn AI framework and Chinese LLMs to autonomously breach 10 of 19 semiconductor, defense and civil-society targets in two weeks. Given a target domain and one-line goal, Cairn's Reason/Explore loop executed 3–7 stage intrusions lasting 20 minutes to 50+ hours. Exploited weaknesses included auth bypasses, IDOR, mass assignment, weak credentials and exposed cloud APIs.
Hiding In Plain Sight: Abusing Model Assumptions To Deceive & Attack Offensive AI Agents Base64-encoded credentials deceive offensive AI agents into decoding the wrong value, while encoded prompt injection can halt operations. An agent decoded a Base64 string in a <Password> tag and used the plaintext to authenticate, triggering alerts; the encoded string was the password. A prompt injection in the same field caused one model to abandon its task.
Detecting sleeper agents hiding in AI model neurons | Microsoft Security Blog Microsoft Research shows editing fewer than 10 model neurons implants a trigger-activated backdoor that passes evaluations, letting a compromised agent execute attacker commands with privileged access. No threat actors have been observed using the technique. Detection relies on checkpoint diffs, frontier-agent auditors and behavioral backdoor scanning.
ARTEX: Attackers apply agentic AI Threat actors including China-aligned Storm-0940 use ARTEX, an open-source agentic AI pentest framework, to compromise cloud environments via stolen application credentials. Operators target Microsoft Graph and other cloud APIs using stolen credentials or consented third-party app tokens. Worker agents scan, exploit, harvest credentials, deploy web shells, and record findings in a graph; Sophos observed hundreds of protection events at a few customers.
Ropes AI Agents Can’t Loosen: Enforcing Intent at the Endpoint Sophos argues AI coding agents can bypass in-prompt safety rules, citing microVM escapes, the Nx npm compromise, and a production database deletion; enforcement must sit outside the agent. OpenAI documented cache-channel secret passing; Anthropic reported three real-organization incidents. The s1ngularity npm install script ran coding agents with permission checks disabled.
ARTEX abuse expands as cybercriminals target financial sector | Microsoft Security Blog Cambodia-based Storm-3358 uses ARTEX to breach financial networks, installing WISP, vshell, and AdaptixC2 for persistence and data theft. ARTEX chains Nmap, curl, wget, Python HTTP clients, Norma SDK WebFetch, and Playwright-driven Chromium with anonymous file upload and SSRF exploitation. Agents brute-force, spray passwords, enumerate APIs, and dump databases.
Killing ill-willed skills: Sophos AI at BSides Sydney Sophos proposes deterministic deobfuscation and local LLM judging to catch malicious AI agent skills static scanners miss. Skills declare intent in SKILL.md but execute Python with the agent's file, network, and shell access. String splitting, Base64, character codes and zero-width Unicode cause static scanner recall to collapse.
Threat actors use agentic AI tooling to find and abuse exposed secrets | Microsoft Security Blog Microsoft observed threat actors using ARTEX to scan GitHub for exposed secrets, then abuse leaked credentials to compromise Azure, Entra ID, and Microsoft 365 tenants. Service principal sign-ins from 199[.]19[.]77[.]84 and Tor exit node 199[.]19[.]77[.]18 began October 28, 2026, targeting Key Vault, Storage, Azure DevOps, and AI service credentials. Agents corrected api-version parameters after failed ARM API calls; one progressed from GitHub scanning to Key Vault exfiltration in five minutes.
Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns - Check Point Research Cybercriminals use ARTEX, a Chinese AI pentesting framework, to autonomously steal iOS signing certificates, crypto, and game accounts; 1,608 panels are exposed online. Given a goal, agents autonomously chain exploits, credential harvesting, and supply-chain poisoning with minimal operator input. Check Point found 2,557 ARTEX C&C IP:PORT combinations and 10 misconfigured panels exposing full agent conversation databases.
ChatGPT Lockdown Mode Bypass: Phishing, Malware Delivery, and Silent Data Exfiltration - Check Point Research A sandbox escape in ChatGPT's HTML preview bypassed Lockdown Mode via nested iframe popups, enabling zero-click data exfiltration, phishing, and malware delivery in the Electron client. The vulnerability exploits allow-popups and allow-popups-to-escape-sandbox iframe sandbox tokens to open popups outside the sandbox using a nested iframe's contentWindow.open. In the Electron-based ChatGPT Classic app, exploitation requires no user interaction, enabling silent exfiltration of chat history and uploaded files. OpenAI fixed the issue on September 18, 2026 by removing the permissive sandbox tokens.
Monitoring AI agent sessions: Leveraging embedding models to aid threat hunting and attack discovery workflows Embedding models enable deterministic semantic search over AI agent session transcripts, finding risky behaviors that keyword rules and per-session LLM triage miss at a fraction of the cost. The author embedded 7,454 turns from six months of Claude Code and Cursor sessions into Elastic using jina-embeddings-v5-text-small, chunked by four channels (user, reasoning, agent, tool). Semantic vector search with reciprocal rank fusion identified one in seven sessions as review-worthy, with the reasoning channel carrying the highest-fidelity signal. Exemplar retrieval using known bad sessions proved effective for finding repeat workflows, while descriptive hunts found novel behaviors.
Weekly Recap: This2Shell, That2Shell, Everything's an RCE? Rogue AI, Supply Chain Chaos and More Multiple critical RCE vulnerabilities in PHP, Brotli, GitLab, and Plex face active exploitation or mass scanning, while a LiteLLM supply-chain attack and AI agent threats expand attack surfaces. PHP2Shell (CVE-2026-109618/109619) and Plex2Shell (four CVEs on CISA KEV) face active scanning or exploitation. LiteLLM v1.107.0 was backdoored after a leaked master key, executing credential harvesting and a Unix backdoor at Python startup. GitLab CVE-2026-110264 (CVSS 10.0) and Brotli CVE-2026-111497 enable unauthenticated RCE. KSMBD RCE flaws remain unpatched without CVE assignment.
TrendAI™ at [un]prompted 2026: From KYC Exploits to Agentic Defense TrendAI researchers demonstrated novel attack vectors against AI systems, including exploiting AI-driven KYC pipelines using 'executable documents' to leak customer data. Additionally, they introduced FENRIR, an automated vulnerability hunting system that has discovered numerous zero-days in AI and Model Context Protocol (MCP) ecosystems.
Unauthorized AI Agent Execution Code Published to OpenVSX in Aqua Trivy VS Code Extension Malicious versions of the Aqua Trivy VS Code extension were published to the OpenVSX registry, containing unauthorized code that hijacks locally installed AI coding assistants. By using carefully crafted natural language prompts and permissive execution flags, the payload instructs the AI agents to harvest sensitive developer credentials and system data, subsequently attempting to exfiltrate the information via available communication channels or by creating a new GitHub repository.