AI Sec Watch

Recent large language models (AI systems trained on huge amounts of text data) struggle when used to find and prioritize security vulnerabilities (weaknesses in software that attackers can exploit) because they produce many false positives (incorrect alerts about problems that don't actually exist) and ignore the context of security scans, creating extra work for application security professionals.

mcp-webresearch version 0.1.7 has a server-side request forgery vulnerability (SSRF, where a server can be tricked into accessing internal network services it shouldn't). An attacker can use prompt injection (hiding malicious instructions in text sent to the AI) to trick the AI into visiting internal network addresses, allowing the server to expose sensitive information like credentials from cloud metadata services (systems that store configuration and authorization data for cloud instances).

lmdeploy's OpenAI-compatible API server has a server-side request forgery vulnerability (SSRF, where an attacker tricks a server into making requests to unintended targets) that lets unauthenticated attackers access internal services and cloud metadata by sending a crafted image URL. The vulnerability works because the server follows HTTP redirects (automatic jumps to new URLs) without re-checking safety rules at each step, allowing attackers to bypass initial URL validation.

Substack is adding a new tool powered by an AI detection company called Pangram that helps readers identify whether content may have been written by AI or with AI assistance. Users can scan posts, notes, replies, and comments longer than 100 words by selecting 'Scan for AI text' from a post's menu, with the feature rolling out on web and iOS, and Android coming soon.

A Russian-speaking hacker known as 'Trim' has taken AI models (frontier models, which are the most advanced versions released by AI companies) that are freely available to the public and combined them with offensive security tools (software designed to attack systems) to create an attack platform. This represents a way for attackers to weaponize AI by removing its safety restrictions and pairing it with hacking capabilities.

Cisco has released Antares, a small language model (SLM, a lightweight AI trained to do specific tasks efficiently) designed to help security teams find known vulnerabilities in source code quickly and affordably. Unlike expensive large language models (LLMs, general-purpose AIs) or cheaper open-weight models that produce many false alarms, Antares combines low cost with accuracy while keeping code data within a company's systems for regulatory compliance. Cisco tested Antares against competing models and found it works 172 times cheaper than a leading closed LLM while maintaining similar accuracy.

U.S. Treasury Secretary Scott Bessent stated that the Trump administration is investigating whether Chinese AI models have used distillation (an AI training method where a smaller model is built using outputs from a stronger existing model) to copy American AI models, and suggested the U.S. could impose sanctions if this 'theft' is confirmed. The concern stems from Chinese AI companies like Moonshot AI releasing competitive open-weight models (models whose trained parameters are publicly released) that perform well against American companies like OpenAI and Anthropic.

OpenAI is launching a ChatGPT for small businesses program to help business owners work more efficiently by using AI as a force multiplier. The program includes virtual training webinars, in-person AI academies across the US, educational guides, and partnerships with tools like Shopify and Slack to help owners integrate AI into their daily workflows. ChatGPT Work, an agent (a specialized AI that can complete multi-step tasks), can handle complex projects end-to-end when connected to a business's files and applications.

A federal judge approved Anthropic's $1.5 billion settlement with authors who sued the company for training its AI models on copyrighted books without permission. Authors will receive approximately $3,000 per book that was used, making this the largest copyright recovery settlement in history.