AI Sec Watch

AI systems change continuously between deployments (such as when retrieval indexes update overnight or new tools are added), which breaks the traditional governance model where compliance is checked after development is complete. Most organizations still treat governance as a separate review layer rather than embedding it into the actual deployment process, leaving companies blind to changes most likely to affect the system. Chinese AI companies instead treat governance as release infrastructure, embedding compliance checkpoints directly into the deployment pipeline so that no product launches without passing these checks.

Anthropic's Project Glasswing, which uses Claude Mythos Preview (an AI model trained to find software bugs), has discovered approximately 10,000 critical or high-severity vulnerabilities across over 1,000 open-source projects and 50 partner organizations. While the AI successfully identified thousands of real vulnerabilities, maintainers are overwhelmed by the flood of bug reports and lack the capacity to patch them quickly, creating a major cybersecurity challenge where finding bugs is now much easier than fixing them.

Anthropic is preparing to publicly release its Mythos model, an advanced AI designed for code analysis that can automatically develop professional-level cyberattacks but also help find and fix software bugs before they're exploited. The company initially delayed public release due to security concerns, but has since developed a guardrail system (safety restrictions built into the model) and is now testing Mythos in Claude Code and Claude Security. Anthropic is also running a project called Glasswing, which partners with companies to use Mythos Preview to find vulnerabilities in critical software, having uncovered 10,000 high or critical-severity vulnerabilities in its first month.

Pope Leo XIV released a papal encyclical (an official open letter from the Catholic Church) called 'Magnifica Humanitas' warning about risks from AI and rapid technological advancement, including AI-powered warfare and job displacement. The document emphasizes that current legal and ethical protections are inadequate to safeguard human dignity as AI adoption accelerates.

This paper presents a method for performing skyline optimization (a technique that filters data to find the most important records based on multiple criteria) on encrypted data that is split across multiple locations in a vertical data federation (a system where different organizations each hold different columns of the same dataset). The researchers developed an asynchronous structured skyline predicate that improves both efficiency and security while protecting sensitive data from unauthorized access.

Researchers developed a new method to attack deep neural networks that analyze 3D point clouds (collections of data points representing 3D objects) by using cage-based deformation, which smoothly warps the entire shape rather than moving individual points. The method generates adversarial attacks (malicious inputs designed to fool AI systems) that look natural to humans while successfully tricking classifiers, and these attacks remain effective even against defense methods.

Federated Learning (FL, a technique where multiple computers train an AI model together without sharing raw data) faces security challenges from adversarial attacks (attempts to trick the model with carefully crafted inputs) and data heterogeneity (when each computer has different types of data). The paper introduces Fed-CDP (Federated Contrastive Diffusion Prototypes), a new approach that uses a server to actively synthesize improved features from client data rather than just collecting them, which helps make the shared model more robust against attacks and reduces model drift (when local models diverge from each other).

This research paper studies how to estimate discrete distributions (collections of data categories and their frequencies) while protecting sensitive information using utility-optimized local differential privacy (ULDP, a privacy protection method that keeps data private locally while allowing more accurate results for non-sensitive information). The authors mathematically prove the fundamental limits of this privacy-utility trade-off and propose new optimal mechanisms called utility-optimized block design schemes to achieve the best possible accuracy under these privacy constraints.

This paper presents MGRNet, a graph reasoning model (a method that uses network structures to understand relationships between data points) designed to improve multi-modal object re-identification (ReID, the task of matching the same object across different image types like visible and infrared photos). The approach handles low-quality local features by constructing modality-aware graphs (structures that represent relationships between image patches while accounting for different image types) and selectively swapping graph nodes to combine local and global information, ultimately creating more reliable object representations.