AI Sec Watch

vLLM (a system for running large language models) versions 0.7.0 through 0.8.x have a bug in how they create hash values (fingerprints) for images. The hashing method only looks at the raw pixel data and ignores important image properties like width and height, so two different-sized images with the same pixels would create identical hash values. This can cause the system to incorrectly reuse cached results or expose data it shouldn't.

vLLM, an inference and serving engine for large language models, had a vulnerability in versions before 0.9.0 where timing differences in the PageAttention mechanism (a feature that speeds up processing by reusing matching text chunks) were large enough that attackers could detect and exploit them. This type of attack is called a timing side-channel attack, where an attacker learns information by measuring how long operations take.

A vulnerability (CVE-2025-5320) was found in Gradio, a web framework for building AI demos, affecting versions up to 5.29.1. An attacker could manipulate the localhost_aliases parameter in the CORS Handler (the component that controls which websites can access the application) to gain elevated privileges, though executing this attack is difficult and requires remote access.

This article collection discusses security challenges in AI and cloud systems, particularly focusing on agentic AI (AI systems that can take autonomous actions). Key risks include jailbreaks (tricking AI systems into ignoring safety rules), prompt injection (hidden malicious instructions in AI inputs), and tool misuse by autonomous agents, which require contextual red teaming (security testing designed for specific use cases) rather than generic testing to identify real vulnerabilities.

Google released Veo 3, a frontier video generation model (an advanced AI system at the cutting edge of technology) that generates both video and audio with high quality and appears to be a marked improvement over existing systems. The model performs well on human preference benchmarks and may represent the point where video generation becomes genuinely useful rather than just a novelty. Additionally, Google announced several other AI improvements at its I/O 2025 conference, including Gemini 2.5 Pro and enhanced reasoning capabilities, while Anthropic released Claude Opus 4 and Claude Sonnet 4 with frontier-level performance.

CVE-2025-5277 is a command injection vulnerability (a flaw where an attacker can trick a program into running unwanted commands) in aws-mcp-server, an MCP server (a software tool that helps AI systems interact with AWS cloud services). An attacker can craft a malicious prompt that, when accessed by an MCP client (a program that connects to the server), executes arbitrary commands on the host system, with a critical severity rating of 9.4.

ClickFix is a social engineering technique (a method that tricks people rather than exploiting technical vulnerabilities) that adversaries are adapting to attack computer-use agents (AI systems that can control computers by clicking and typing). The attack works by deceiving users into believing something is broken or needs verification, then tricking them into clicking buttons or running commands that compromise their system.

This article describes a curated database of AI literacy training programs across Europe designed to help organizations and professionals comply with Article 4 of the EU AI Act (a regulation requiring organizations to build employee understanding of AI). The programs are selected based on whether they teach what AI is, its risks and benefits, and how to use it responsibly in the workplace.

This content discusses security challenges in agentic AI (autonomous AI systems that can take actions independently), emphasizing that traditional jailbreak testing (attempts to trick AI into breaking its rules) misses real operational risks like tool misuse and data theft. The material suggests that contextual red teaming (security testing that simulates realistic attack scenarios in specific business environments) is needed to properly assess vulnerabilities in autonomous AI systems.