Introduction

The popular open-source AI gateway LiteLLM was recently hit by a supply chain attack, with a malicious version automatically stealing API keys for cloud and AI services. Today, besides walking you through the incident response guide, I will also discuss the attack planning capabilities of cutting-edge AI models in industrial control systems. As for how a major cybersecurity firm built a defensive model at just one-two-hundredth of the training cost of a general-purpose model, I will share that with you shortly.

Today's AI Security Highlights

1. Open-Source AI Gateway LiteLLM Hit by Supply Chain Poisoning, Malicious Version Implants Remote Access Tool

• What Happened: According to the Microsoft Security Blog, version 1.107.0 of the open-source AI gateway LiteLLM was compromised via supply chain poisoning on October 21, 2026. Attackers used stolen GitHub credentials to publish packages and Docker images containing malicious code. This malicious version automatically executes when Python starts, stealing API keys for cloud, browser, and AI providers, and includes a remote access trojan (RAT).
• Impact: Developers and enterprises using LiteLLM version 1.107.0 face the risk of credential leakage and remote control of their cloud infrastructure and AI services.
• Mitigation: Affected users should immediately delete the relevant malicious files, downgrade or rebuild to version 1.106.2, and comprehensively rotate any potentially compromised credentials.
• Source: Microsoft Security Blog (https://www.microsoft.com/en-us/security/blog/)

2. Industrial Systems Security Report: Cutting-Edge AI Achieves Up to 90% Success Rate in Attack Planning

• What Happened: Cybersecurity firm Frenos released the "State of AI in OT Security" report, evaluating the performance of cutting-edge and open-source AI models across more than 500 simulated Industrial Control System (ICS) attack scenarios. The study shows that the three most powerful cutting-edge models could plan viable attack paths in approximately 90% of attempts, though they still rely heavily on guidance and context from human experts.
• Impact: Critical infrastructure operators, such as managers of industrial control systems in energy, manufacturing, and water resources, need to re-evaluate the potential threats AI poses to physical security defenses.
• Mitigation: It is recommended that enterprises prioritize strengthening detection and response mechanisms during a two-year preparation window, and leverage AI to assess their systems' overall exposure.
• Source: Industrial Cyber (https://industrialcyber.co/)

3. ESET Launches Dedicated Cybersecurity AI Model NeoCortex, Featuring Low Cost and Guardrails Against Rogue Behavior

• What Happened: ESET launched "ESET NeoCortex," a foundational AI model designed specifically for cybersecurity, to power its LiveCortex correlation and response engine. Through highly optimized training, this model delivers precise defensive analysis at just one-two-hundredth of the training cost of general-purpose AI models, and features built-in strict guardrails to prevent AI agents from going rogue.
• Impact: Enterprises adopting the ESET LiveCortex correlation and response engine can obtain secure-by-design AI defensive analysis services at a much lower computational cost.
• Mitigation: When deploying AI agents, enterprises should reference built-in safety mechanisms designed to prevent rogue behavior, such as restricted sandboxes, read-only tools, token limits, and emergency kill switches.
• Source: ESET (https://www.eset.com/)

4. Oracle Partners with OpenAI to Launch Security Mesh, Strengthening Enterprise AI Infrastructure Defense

• What Happened: Oracle announced the launch of "Oracle AI Security Mesh," combining OpenAI's cutting-edge cybersecurity models with Oracle's testing tools and customer system context. This service proactively finds and patches vulnerabilities by analyzing application behavior, complementing OCI's Zero Trust Packet Routing and identity verification capabilities.
• Impact: Enterprises running autonomous AI agents or AI workloads on Oracle Cloud Infrastructure (OCI).
• Mitigation: When deploying autonomous AI agents, enterprises should establish infrastructure-level security perimeters to prevent compromised workloads from lateral movement or data exfiltration.
• Source: Oracle (https://www.oracle.com/)

5. US Mortgage Giant Fannie Mae Updates Cybersecurity Rules, Requiring Incident Reporting Within 36 Hours

• What Happened: The Federal National Mortgage Association (Fannie Mae) released an update to its servicing guide, renaming all Artificial Intelligence (AI) terminology to "Super Intelligence (SI)." The new guide introduces a strict 36-hour cybersecurity incident reporting window and adds several requirements for the development and use of SI systems, including establishing SI risk assessments, data governance standards, and model training controls.
• Impact: Fannie Mae's partners, third-party vendors, and financial institutions and mortgage servicers using its systems.
• Mitigation: Enterprises partnering with financial institutions should immediately review their third-party vendor compliance and establish cybersecurity incident response processes capable of reporting within 36 hours.
• Source: American Bankers Association (https://www.aba.com/)

Conclusion

The supply chain vulnerability in the popular open-source gateway LiteLLM highlights that AI infrastructure has become a primary target for hackers. I recommend that you audit the versions of AI packages used by your team today to ensure you are not using any compromised images. I'm Muyan, and I'll see you tomorrow.