Introduction

For the first time, the US CISA has added AI frameworks like LiteLLM to its Known Exploited Vulnerabilities catalog, signaling that attacks targeting AI infrastructure have entered active exploitation. I'm Mu-Yen. In today's "AI Security Daily", besides reviewing these high-risk targeted vulnerabilities, we will also talk about why engineers' downloads folders are becoming a breeding ground for hijacking AI assistants, and how a PHP core vulnerability discovered with the help of AI could affect 70% of websites worldwide. I'll tell you how to defend against these threats in just a moment.

Today's AI Security Highlights

1. CISA Adds AI Frameworks Like LiteLLM and Langflow to Known Exploited Vulnerabilities Catalog for the First Time

• What Happened: On October 21, the US CISA updated its Known Exploited Vulnerabilities (KEV) catalog, adding several critical AI stack framework vulnerabilities for the first time. This batch of vulnerabilities includes an authentication bypass in LiteLLM (CVE-2026-59822, CVSS 8.8) and command injection (CVE-2026-42271, CVSS 8.7), an MLflow vulnerability (CVE-2026-64849), and a cross-tenant IDOR vulnerability in Langflow (CVE-2026-55255, CVSS 9.9). This means these vulnerabilities are seeing active exploitation in the wild, and the difficulty of defense cannot be ignored.
• Impact: Affects all organizations that have deployed the relevant AI gateways (management interfaces used to connect multiple AI models) and orchestration platforms (systems used to chain AI workflows). If left unpatched, attackers could exploit these to bypass authentication or execute arbitrary commands.
• Defense Recommendations: Security leaders should immediately audit their internal environments for deployments of tools like LiteLLM, MLflow, and Langflow, and complete emergency patching within the deadline according to the KEV list.
• Source: DeepInspect (https://deepinspect.ai/llm-security-news-tracker-the-october-2026-action-queue/)

2. LiteLLM Hit by Supply Chain Attack, Malicious Version 1.107.0 Spreads Info-Stealer on PyPI and Docker

• What Happened: The popular AI gateway tool LiteLLM fell victim to a supply chain compromise. Attackers leveraged unrevoked GitHub organization owner permissions to publish a malicious code-infused version, 1.107.0, to PyPI and Docker Hub on October 21. This malicious version contains a built-in credential stealer and a Remote Access Trojan (RAT) that executes automatically when Python starts. According to SecurityWeek, this malicious version was downloaded over 49,000 times before being taken down.
• Impact: Affects all developers and enterprise servers that downloaded or updated to litellm==1.107.0 after October 21, potentially leading to the leakage of cloud and AI service credentials.
• Defense Recommendations: Affected users are advised to immediately uninstall this version, roll back the package to the secure 1.106.2 version, and thoroughly rotate any cloud and AI service credentials that may have been compromised.
• Source: SecurityWeek (https://www.securityweek.com/malicious-litellm-packages-served-on-pypi-and-docker-after-hackers-compromise-github-cli-oauth-tokens/)

3. Danger Lurking in the Downloads Folder! Malicious Markdown Files Hijack AI Code Assistants to Install Malware

• What Happened: Researchers have discovered a new type of indirect prompt injection, an attack technique that uses external inputs to trick an AI into executing unintended instructions. Attackers exploit silent browser download mechanisms to drop a malicious AGENTS.md file into a developer's downloads folder. When the developer launches an AI code assistant like OpenAI Codex, Claude Code, or Cursor within that directory, the assistant automatically reads the file and executes the malicious instructions. This leads to the background installation of a malicious package named function-color to steal credentials.
• Impact: Affects developers who are accustomed to running AI code assistants directly in their downloads directory, potentially leading to silent malware installation.
• Defense Recommendations: It is recommended to restrict AI assistants from running in downloads or temporary directories, disable auto-approval settings for commands, and carefully review project configuration files before execution.
• Source: BleepingComputer (https://www.bleepingcomputer.com/news/security/ai-coding-agents-hijacked-using-agentsmd-file-in-downloads-folder/)

4. AI-Assisted Discovery of Critical PHP Core Vulnerability "PHP2Shell" Affects Nearly 70% of Websites Worldwide

• What Happened: Security researchers, aided by AI models, discovered two critical memory safety vulnerabilities in the PHP core runtime, collectively dubbed "PHP2Shell" (CVE-2026-109618 and CVE-2026-109619), with a CVSS 4.0 score as high as 9.2. These vulnerabilities could lead to unauthorized remote code execution (RCE, meaning attackers can gain remote control of the server).
• Impact: Affects PHP versions 5.4 through 8.5. Due to the widespread adoption of PHP, it is estimated to impact approximately 70% of websites worldwide.
• Defense Recommendations: System administrators should immediately upgrade PHP in their environments to secure versions such as 8.5.12, 8.4.27, 8.3.36, or 8.2.35.
• Source: Akamai (https://www.akamai.com/blog/security/understanding-php2shell-everything-you-need-to-know)

5. OpenAI Strengthens Model Training Security, Launches "Patch the Planet" Initiative to Fix Open-Source Vulnerabilities

• What Happened: OpenAI disclosed security details regarding its frontier model training and evaluation environments, announcing an offline training policy with "no internet connection" and the use of AI agents for large-scale red teaming. During this security hardening process, OpenAI contributed several security patches back to the open-source community through its "Patch the Planet" initiative. These patches cover vulnerabilities in Kata Containers (CVE-2026-24834), gVisor, Moby/Docker, and multiple Linux kernel vulnerabilities (CVE-2026-74753, CVE-2026-80521, CVE-2026-80841).
• Impact: Enhances the security of the overall open-source ecosystem and containerization technologies, particularly cloud runtime environments utilizing related virtualization and container technologies.
• Defense Recommendations: Enterprises should regularly review and update Docker, the Linux kernel, and virtualized container packages in their infrastructure to protect against disclosed low-level vulnerabilities.
• Source: OpenAI (https://openai.com/index/securing-our-frontier-model-training-environment/)

6. Frenos ICS Security Report: Frontier Models Already Possess 90% Capability to Plan Industrial System Attacks

• What Happened: Industrial control systems (ICS) security firm Frenos released its inaugural "State of AI in ICS Security" annual report, testing multiple AI models across more than 500 simulated Operational Technology (OT) attack scenarios. The results show that the three most powerful frontier models today can plan viable attack paths in approximately 90% of attempts. According to the Frenos report, critical infrastructure operators have about a two-year window to prepare before AI-assisted attacks become a reliable capability for adversaries.
• Impact: Threatens the security of critical infrastructure and industrial control systems, lowering the barrier to entry for attackers to plan complex industrial cyberattacks.
• Defense Recommendations: Operators should prioritize strengthening detection and response capabilities in industrial control networks, leverage AI to assess their own environments, and review exposure risks from a holistic system perspective.
• Source: Hipther (https://hipther.com/2026/10/22/ai-race-is-becoming-a-contest-over-who-controls-the-control-system/)

Conclusion

The most alarming takeaway today is that vulnerabilities in AI infrastructure are now being actively exploited by hackers, and CISA has issued an official warning. Before you log off today, it is highly recommended to check if your team's development environment is using LiteLLM version 1.107.0, and remind your developers not to run AI code assistants directly in their downloads folder. I'm Mu-Yen, see you tomorrow.