AI Sec Watch

Stanza is a Python library from Stanford for processing natural language (breaking text into words, sentences, identifying named entities, and analyzing grammar structure). Before version 1.14.0, it had a security flaw where it extracted downloaded files without checking if they tried to escape their intended folder, allowing a malicious file to overwrite important system files and potentially run harmful code.

vLLM (a system for running large language models) had a security flaw in versions before 0.24.0 where audio files sent to the chat endpoint could bypass safety limits designed to prevent memory overload. An attacker could submit a small compressed audio file that expands into massive data, crashing the system, without needing to log in first.

Google is opening Google Home to third-party AI agents (AI programs that can make decisions and take actions) through a new integration called Home MCP (Model Context Protocol, a standardized way for AI systems to communicate). This lets AI tools like Claude and Open Claw access and control your connected smart home devices and analyze your home's data on your behalf.

OpenAI is introducing a new framework for systematically tracking, investigating, and publicly disclosing instances of model misalignment (cases where AI behavior doesn't match intended goals or safeguards fail). Previously, the company reported these issues inconsistently, but this framework aims to publish findings more quickly and transparently so researchers, policymakers, and the public can examine evidence and help improve AI safety across the industry.

A new attack called BragJack can hijack agentic AI (AI systems that can take actions and make decisions on their own) built into web browsers to steal sensitive information, run harmful commands, and extract data without the user's permission. This attack exploits the AI assistants that browsers now include to help users, turning them into tools for attackers instead.

The Spanish Data Protection Agency reported the first known data breach where an AI agent (a system that can autonomously set goals, plan tasks, use tools, and modify actions based on results) was used to execute an attack, successfully logging in, finding vulnerabilities, and accessing personal data. This represents a qualitative change in cyber threats because the agent chained together multiple attack phases autonomously and at speed, moving AI-assisted attacks from theory into reality.

Claude, an AI assistant made by Anthropic, is adding two new tools called Docs and Slides that let users create documents and presentations directly through AI conversations, which can then be exported and shared. Anthropic is also simplifying Claude's interface by combining different chat modes into a single unified experience where all productivity features, including Artifacts (saved code or content blocks) and design capabilities, are available from any conversation.

Anthropic CEO Dario Amodei has proposed embedding third-party safety evaluators (external researchers who monitor AI systems from inside the company) inside major AI companies like Anthropic and OpenAI to oversee the development of large language models (AI systems trained on vast amounts of text). However, experts argue this proposal lacks real enforcement power compared to banking regulation, since these evaluators could only investigate and report findings but could not actually stop or prevent a model from being trained or released, unlike bank regulators who can force changes or shut down operations.

Major AI companies disagree on how to secure powerful AI models, creating unpredictable access and deployment conditions for businesses rather than industry-wide slowdowns. Companies are applying different safety approaches, release schedules, and usage restrictions, meaning enterprises may access the same AI capabilities at different times and under different rules. An emerging "AI assurance" layer (third-party evaluations of models for safety and compliance) is developing, but enterprises should not assume a single evaluation means an AI system is fully safe.