AI Sec Watch

LibreChat, a ChatGPT-like application supporting multiple AI providers, has an IDOR vulnerability (insecure direct object reference, where an attacker can access or modify resources belonging to other users) in its API key management system in versions up to 0.7.6. An authenticated attacker can inject a userId parameter to overwrite another user's API keys, potentially stealing their API key configurations or blocking their service.

OpenAI has upgraded GPT-5.5 Instant to respond more accurately with more natural, human-like language and shorter responses, while retiring older models like o3 (August 26) and GPT-4.5 (June 27) to focus resources on newer capabilities. The company is also adding a job search tool to ChatGPT that integrates with job boards like Indeed and Upwork to help users find positions and tailor resumes.

Microsoft announced two new text-based LLMs (large language models, AI systems trained on text data): MAI-Thinking-1, a 35-billion-parameter model (parameters are the adjustable values that make up a neural network) designed for reasoning tasks, and MAI-Code-1-Flash, a smaller 5-billion-parameter model built specifically for code generation in GitHub Copilot and VS Code. Both models were trained on licensed data rather than web scrapes, and Microsoft claims MAI-Thinking-1 outperforms Claude's Sonnet 4.6 model despite its smaller size.

HP released patches for a critical buffer overflow vulnerability (a coding flaw where too much data is written into a fixed-size memory container) in its Poly Voice conference phones that could allow attackers without authentication to gain root access (complete control of the operating system) and record conversations for voice deepfakes (AI-generated fake audio impersonations). The flaw exists in code that processes ICE (Interactive Connectivity Establishment, a feature for establishing direct network connections) requests and affects multiple Poly phone models.

Microsoft announced new AI models including MAI-Code-1-Flash (a model that generates source code from written descriptions) and MAI-Thinking-1 (a reasoning model) to reduce dependence on OpenAI and lower costs for developers. These models run on Microsoft's own Azure cloud infrastructure, allowing the company to avoid paying third parties while offering developers lower token costs (tokens are the basic units that an AI model reads and processes). Microsoft is positioning itself to compete directly with proprietary models from OpenAI and Google by building its own AI capabilities across multiple layers of the technology stack.

President Trump signed an executive order establishing a voluntary framework for the federal government to review the national security risks of the most advanced AI systems (models built by companies like OpenAI and Google that represent the cutting edge of AI development) before their public release, with a 30-day review period. The order aims to balance security concerns with concerns about slowing innovation, and it allows frontier labs to voluntarily share their most advanced models to help secure critical infrastructure and strengthen government cyber defenses.

Hackers reportedly used Meta's AI support chatbot to take over high-profile Instagram accounts by tricking it into changing email addresses linked to those accounts, affecting pages associated with the Obama White House, Sephora, and other notable figures. The incident wasn't simply a case of prompt injection (tricking an AI by hiding instructions in its input), but rather revealed a deeper problem with how the AI chatbot was designed to handle account recovery requests.

Microsoft held its Build 2026 conference with announcements from CEO Satya Nadella covering new hardware and AI updates, including a Surface RTX Spark Dev Box designed to help developers run local AI models (machine learning systems that operate on a user's own device rather than in the cloud) on their computers. The event also featured updates to Microsoft's own AI models and an always-on personal assistant feature.

AI agents (systems that can act independently to complete tasks) with high autonomy and broad permissions are very difficult to secure and pose a serious risk to enterprises. The article warns that companies need to take action now to prevent AI agents from causing major problems in the future.