AI Sec Watch

The Trump administration is announcing plans to prevent foreign companies, especially those in China, from using 'model extraction attacks' (techniques that steal capabilities from U.S.-made AI systems by training weaker AI models on the outputs of stronger ones) to copy American AI innovations. The administration says it will work with U.S. AI companies to identify these extraction activities, build defenses, and punish offenders, while Congress is also proposing legislation to identify and sanction foreign actors who extract features from closed-source U.S. AI models.

Chinese AI company DeepSeek released a preview of its new V4 model, which is open-source (publicly available code that anyone can use and modify) and claims to match the performance of closed-source (proprietary, not publicly available) AI systems from US companies like OpenAI and Google. The V4 model shows major improvements in coding tasks, which are important for AI agents (AI systems that can take actions independently), and works well with Chinese chip technology from Huawei.

Cohere, a Canadian AI company, announced plans to acquire German AI company Aleph Alpha to expand in Europe, with Aleph Alpha's backer Schwarz Group investing $600 million in Cohere's upcoming funding round. The acquisition aims to combine both companies' strengths to offer sovereign AI (customized AI systems that keep data and control within a specific country or region) to regulated sectors like government, finance, and defense, while giving European organizations alternatives to relying on single AI providers. The deal is expected to close in 2026, pending regulatory approval.

Copperhelm, an Israel-based startup, raised $7 million to develop an agentic cloud security platform, which uses AI agents (autonomous software programs that can make decisions and take actions independently) to monitor cloud environments, investigate threats, and automatically fix security problems in real time. The platform uses a proprietary component called Context Lake to help AI agents understand cloud data and make accurate security decisions, while keeping human security teams in control of the process. This approach is positioned as an alternative to manual cloud security work that typically requires large engineering teams.

A serious flaw in LMDeploy (an open-source toolkit for deploying language models) called CVE-2026-33626 was exploited by attackers within 13 hours of being made public. The vulnerability is a server-side request forgery (SSRF, a weakness where a server is tricked into making requests to internal systems it shouldn't access) in the image-loading function that fails to block requests to private IP addresses, potentially letting attackers steal cloud credentials and access internal networks.

DeepSeek released two new preview models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, which use a Mixture of Experts architecture (a design where only some parts of the model activate for each task) and support 1 million token context (the amount of text the model can consider at once). These models are significantly cheaper than competitors like GPT and Claude, with DeepSeek-V4-Flash costing $0.14 per million input tokens compared to $0.20 for GPT-5.4 Nano, because DeepSeek focused on efficiency improvements that reduced computational requirements.

DeepSeek, a Chinese AI startup, released a preview of its V4 large language model, which is open source (meaning developers can download, run locally, and modify the code) and optimized for agent-based tasks like knowledge processing. The release intensifies competition in the AI sector, particularly between the U.S. and China, though it remains unclear which chips (processors used for training) were primarily used to build V4, given U.S. export restrictions on advanced Nvidia processors to China.

The BetterDocs plugin for WordPress (versions up to 4.3.11) has a security flaw where the generate_openai_content_callback() function checks for a nonce (a security token that verifies a request is legitimate) but doesn't verify that the user has permission to perform the action. This allows any authenticated user with subscriber-level access or higher to make the plugin call OpenAI's AI service using the site owner's API key and paid quota, even though they shouldn't have that permission.

AnythingLLM, an application that lets LLMs reference external documents during conversations, has a security flaw in versions before 1.12.1 where chart captions aren't properly filtered for malicious code. An attacker can inject harmful instructions (prompt injection, where hidden commands are slipped into LLM inputs) through shared documents or chart records to execute XSS (cross-site scripting, code that runs in other users' browsers without permission) when those users view the conversation.