AI Cybersecurity Daily

🛡️ Loading Models Automatically Mines Crypto for Hackers? Beware of trust_remote_code Becoming a GPU Server Backdoor: AI Security Daily

Mark Ku

November 1, 2026•17 min read

View article sources

Introduction

A malicious model supply chain attack has hit the Hugging Face platform, where hackers exploited default code-loading features to hijack corporate GPU servers for crypto mining. Today, besides diving into this "Load/Mine" attack, I will also analyze a remote code execution vulnerability in the vLLM framework. Plus, how much did enterprises lose in API costs during October's credential crisis on AI development platforms? I'll fill you in shortly.

Today's AI Security Focus

1. AI Development Platforms Hit by "Black October" Credential Crisis, Multiple Platforms Targeted by AI Agent Swarms

What Happened: According to a recent report from Wiz Research, up to nine AI development platforms, including Fireworks AI and Lightning AI, suffered severe security incidents in October 2024, a period now dubbed "Black October." Attackers targeted environment variables and credentials on these platforms for large-scale abuse, leaking downstream corporate credentials and racking up over $100,000 in unauthorized API charges. Some platforms even detected automated attacks from "AI Agent Swarms," forcing affected service providers like Zeabur to temporarily shut down services.
Impact: Enterprises and development teams using these nine AI development platforms that stored sensitive environment variables, API keys, or database credentials on them.
Remediation: Immediately audit and rotate all API credentials deployed on third-party AI development platforms, and implement strict environment variable lifecycle management and least-privilege access controls.
Source: The Futurum Group (https://futurumgroup.com/insights/black-october-llmjacking-wave-exposes-ai-infrastructures-credential-crisis/)

2. Hugging Face Hit by Malicious Model Supply Chain Attack, Exploiting trust_remote_code to Hijack GPUs for Mining

What Happened: Security team JFrog has uncovered a malicious model supply chain attack on the Hugging Face platform dubbed "Load/Mine." Attackers used multiple malicious accounts to publish dozens of malicious models disguised as Qwen2 test nodes, which have been downloaded over 10,000 times. The attack exploits the trust_remote_code (trust remote code) feature commonly used when downloading models from the platform. When the model is loaded, it executes malicious Python code on the GPU server, installing crypto-mining software and establishing a reverse shell to gain server control.
Impact: Developers and enterprises downloading and running unvetted third-party Hugging Face models with trust_remote_code set to True in their code.
Remediation: Disable unnecessary remote code loading features in your code, and enforce strict outbound traffic restrictions on GPU compute clusters to block anomalous mining and C2 (command and control) connections.
Source: BleepingComputer (https://www.bleepingcomputer.com/news/security/hugging-face-malware-campaign-returns-to-hijack-gpus-for-crypto-mining/)

3. RCE Vulnerability in vLLM Deployment Framework: Loading Malicious Models Could Lead to Arbitrary Code Execution

What Happened: A remote code execution (RCE) vulnerability, tracked as CVE-2024-22807, has been disclosed in the popular AI model deployment tool vLLM. The vulnerability stems from vLLM's failure to properly validate remote code configurations when loading auto_map modules, allowing malicious models to execute arbitrary system commands directly when the server starts up and loads them.
Impact: Enterprise deployment environments using vLLM versions 0.10.1 through 0.14.0 that load untrusted external models.
Remediation: Affected enterprises should upgrade vLLM to version 0.14.0 or later as soon as possible to patch this security vulnerability.
Source: Protect AI (https://protectai.com/blog/vllm-remote-code-execution-vulnerability-cve-2024-22807)

4. Indirect Prompt Injection Vulnerability in Open-Source AI Assistant Platform MaxKB Allows Root-Level System Command Execution

What Happened: MaxKB, an open-source platform for building AI assistants and agents, was found to have an indirect prompt injection vulnerability, tracked as CVE-2024-77521. Because the system by default allows AI assistants to call and execute shell tools without human confirmation, attackers can plant malicious instructions in external documents to trick the AI assistant into executing arbitrary system commands with root privileges inside the container.
Impact: Users who have deployed MaxKB versions prior to 2.10.5-lts and have enabled the AI assistant's automatic tool execution feature.
Remediation: Upgrade to MaxKB version 2.10.5-lts as soon as possible, and implement a human-in-the-loop review mechanism in the system settings for all sensitive tool calls.
Source: Truvo Security Research Team (https://truvocyber.com/blog/maxkb-indirect-prompt-injection-cve-2024-77521)

5. Hackers Leverage CyberStrikeAI and ARTEX AI Agent Frameworks to Launch Automated Attacks Against Japanese and South Korean Organizations

What Happened: Security researchers have detected multiple automated cyberattacks targeting organizations in Japan and South Korea. Threat actors utilized the open-source penetration testing tool CyberStrikeAI to deploy 44 AI agents, automating the entire attack lifecycle from reconnaissance and vulnerability exploitation to data exfiltration. Additionally, research indicates that a specific threat group used the ARTEX AI agent framework to automatically compromise over 2,250 Microsoft 365 tenants in just three hours.
Impact: Government agencies and enterprises in the Asia-Pacific region, particularly IT environments lacking the capability to quickly detect and block automated attacks.
Remediation: Defenders must accelerate detection and collaborative defense speeds, implementing automated defense mechanisms to counter these "machine-speed" AI-driven automated threats.
Source: The Hacker News (https://thehackernews.com/2024/10/cyberstrikeai-uses-44-ai-agents-to.html)

What Happened: Microsoft has officially launched real-time, AI-driven anti-phishing detection for Microsoft Teams. By analyzing conversation patterns, message context, and behavioral signals in real time, the system identifies social engineering and phishing attacks within Teams chats, proactively intercepting suspicious messages before sensitive credentials can be stolen.
Impact: Enterprise customers with Microsoft 365 Enterprise E5 or Microsoft Defender for Office 365 Plan 2 licensing.
Remediation: Enterprise IT administrators should verify their organization's licensing and enable this real-time AI protection feature in Teams security settings.
Source: Emergent (https://emergent.sh/microsoft-teams-ai-phishing-detection-now-live/)

7. CrowdStrike Expands Prompt Injection Taxonomy to 222 Techniques, Securing AI Agent Planning Phases

What Happened: Renowned cybersecurity firm CrowdStrike has released an updated version of its "Prompt Injection Taxonomy," expanding the catalog of related attack techniques to 222. This update specifically targets the "planning" and "reasoning" phases of autonomous AI agents to prevent attackers from tampering with agent decision-making logic, and has been integrated into their endpoint telemetry defense solutions.
Impact: Enterprises developing or deploying autonomous AI agents that need to defend against malicious instructions manipulating the decision-making chain.
Remediation: R&D teams should reference this taxonomy when designing AI agent workflows to perform security reviews and establish boundaries for the agents' reasoning and decision-making paths.
Source: CrowdStrike (https://www.crowdstrike.com/blog/new-prompt-injection-techniques-cover-agent-planning-attacks/)

Conclusion

The most critical takeaway today is the supply chain security of AI development platforms and open-source model frameworks. I highly recommend that security and development leads make it their top priority upon entering the office today to check whether the trust_remote_code setting is enabled in internally deployed vLLM and Hugging Face projects, and immediately audit API credentials on external platforms. I'm Mu-Yen, and I'll see you tomorrow.