AI Sec Watch

Meta launched Muse Spark, a new closed-source AI model (a large language model that processes and generates text), marking a shift from its previous open-source Llama models toward a paid subscription approach similar to competitors like OpenAI and Google. While Muse Spark shows competitive performance in text and vision tasks, investors are waiting to see Meta's strategy for driving consumer adoption and generating revenue beyond just improving its advertising business.

This newsletter covers multiple AI developments including a legal battle between Elon Musk and OpenAI's leadership over the company's for-profit status, the gap between AI hype and actual profitability, and the rise of weaponized deepfakes (AI-generated fake videos or images used maliciously) that are spreading misinformation and harming vulnerable groups. The content also reports on business moves like OpenAI ending its exclusive partnership with Microsoft and various regulatory actions worldwide.

Vision-language models (AI systems that process both images and text together) can leak private information from user-uploaded content, such as identifying people in photos or extracting sensitive text. This research examines privacy risks when users submit images and text to these models. The paper proposes privacy-preserving methods to protect user data while still allowing these AI systems to function effectively.

This survey paper examines algorithm debt in machine learning and deep learning systems, which refers to the long-term costs and problems that accumulate when developers use suboptimal algorithms or methods in AI projects. The paper defines what algorithm debt is, identifies warning signs called 'smells' that indicate its presence, and discusses future research directions. Understanding algorithm debt helps developers recognize when quick, temporary solutions in AI projects create technical problems that become harder and more expensive to fix later.

CISOs (chief information security officers) struggle with unpredictable costs when using agentic AI (autonomous AI agents that can make decisions and take actions) for cybersecurity defense, since they are charged per AI token (a unit of text similar to a word) used, and attack volumes can spike unexpectedly. Sevii launched Cyber Swarm Defense, a new mode that charges by protected asset (like laptops or cloud servers) at a fixed yearly rate instead of per token, making defense costs predictable regardless of how many attacks occur. The system also includes governance controls that let security teams automatically remediate low-risk assets while keeping critical ones for human review.

LeRobot, Hugging Face's open-source robotics platform, has a critical unpatched vulnerability (CVE-2026-25874, CVSS score 9.3) that allows unauthenticated attackers to execute arbitrary code by sending malicious data through unencrypted network connections. The flaw stems from unsafe deserialization (a process of converting data back into code without properly checking if it's trustworthy) using pickle, an unsafe data format, which enables attackers to compromise the server, steal sensitive data, or impact connected robots.

Google has reportedly signed a classified agreement allowing the US Department of Defense to use its AI models for 'any lawful government purpose,' despite employee concerns about potential harmful uses. This deal places Google alongside other AI companies like OpenAI and xAI that have made similar classified agreements with the government.

Anthropic announced Claude Mythos Preview, an AI model that can autonomously find and weaponize software vulnerabilities (weaknesses in code that attackers can exploit) without human expert help, though the company is limiting its release to avoid security risks. The announcement highlights how AI capabilities have advanced rapidly over recent years, raising concerns about how cybersecurity defenses can adapt to AI-powered vulnerability discovery.

At DARPA's Artificial Intelligence Cyber Challenge, AI-powered bug-finding systems (automated tools that scan code to detect flaws) successfully identified most artificially inserted vulnerabilities in 54 million lines of code, and notably discovered over a dozen real bugs that weren't part of the test. This demonstrates that AI security tools are becoming increasingly capable at finding both known and unknown vulnerabilities in software.