AI Sec Watch

ChromaDB, a popular vector database used in AI applications, has a critical vulnerability (CVE-2026-45829) that allows unauthenticated attackers to run arbitrary code on servers. The flaw exists because ChromaDB checks authentication after it has already downloaded and executed a malicious model from Hugging Face, meaning attackers can trick the system into running their code by uploading a malicious model and requesting ChromaDB to use it.

Twig (a PHP template engine) has a vulnerability where template names in `{% use %}` tags aren't properly escaped, allowing attackers to inject arbitrary PHP code that executes when the template cache loads. This bypasses Twig's security sandbox, giving attackers remote code execution (the ability to run commands on the server).

LiteLLM versions before 1.83.10 have a vulnerability where users can change their own role to proxy_admin (an administrative role) through the /user/update endpoint, giving them full control over the system including all users, teams, and API keys. Even users with org_admin privileges can exploit this flaw without needing to chain it with other attacks.

LiteLLM versions before 1.83.14 have a privilege escalation vulnerability (a security flaw that lets someone gain higher-level permissions than they should have) where authenticated internal users can create API keys (credentials for accessing the system) that grant access to admin-only routes without proper verification. This allows attackers to bypass role-based access controls (the system that restricts what different users can do) and gain full admin privileges.

This article discusses how security leaders (CISOs, or Chief Information Security Officers) should prepare for AI systems that can take independent actions (agentic AI). The key challenge is creating an AI bill of materials (AI BOM, a detailed list of all components and dependencies in an AI system) that documents both what components make up the AI system and how those components actually behave when running.

This is a discussion panel about how AI companies are working to build systems that understand the physical world, moving beyond the current limitations of LLMs (large language models, which are AI systems trained on text). The conversation explores recent developments in world models, which are AI systems designed to understand and predict how the physical world works.

The `mcp-server-kubernetes` tool had a security flaw where access control settings (environment variables that limit which Kubernetes operations are available) only worked when listing tools, but not when actually running them. This meant an attacker or misconfigured AI agent could bypass these restrictions and run any Kubernetes command, like deleting pods or accessing containers, even if they were supposed to be blocked.

Microsoft is negotiating to supply its custom Maia AI chips to Anthropic, a company that makes Claude, a popular AI assistant. This deal would help Microsoft compete with Amazon and Google in providing specialized AI hardware to clients, while Anthropic seeks to address its computing capacity challenges after experiencing rapid growth in demand for its AI tools.

Amazon SageMaker Python SDK has a vulnerability where it stores an HMAC signing key (a cryptographic secret used to verify that model files haven't been tampered with) in plaintext as an environment variable that can be read by anyone with access to certain AWS APIs. An attacker with the right permissions could steal this key, use it to forge valid model files, and run malicious code on the system running the model.