AI Sec Watch

A developer claims to have reverse-engineered Google DeepMind's SynthID system, which is a watermarking technology that embeds hidden marks in AI-generated images to prove their origin. The developer says they can strip these watermarks from images or add fake ones, though Google disputes this claim.

Asynchronous federated learning (AFL, where multiple devices train a shared AI model without waiting for each other to finish) is faster than synchronous methods but more vulnerable to Byzantine attacks (when some devices send false or corrupted data to sabotage the model). Researchers propose Belisa, a framework that uses feature fingerprints (unique patterns in how local models represent data) to identify and filter out malicious devices, improving robustness and efficiency in real-world scenarios where devices have different data and hardware capabilities.

This research proposes ENClose, a framework that lets control systems (automated systems that adjust themselves based on feedback) operate securely using fully homomorphic encryption, or FHE, a cryptographic method that keeps data encrypted while performing calculations on it. The main innovation addresses two problems: noise building up in encrypted feedback loops and the slowness of doing complex nonlinear operations (calculations that don't follow straight-line relationships) on encrypted data. ENClose uses techniques like function segmentation and tree-based selection to speed up these encrypted calculations by 3 to 20 times compared to previous methods, as demonstrated in real-world tests like vehicle formation control and anomaly recovery.

Large Vision-Language Models (LVLMs, which are AI systems that process both images and text) show promise for general AI but can produce biased outputs, a problem that hasn't been thoroughly studied. Researchers created VLBiasBench, a comprehensive benchmark (a standardized test for measuring performance) that evaluates nine types of social biases in these models, including age, gender, race, and disability status, using 128,342 test samples generated with images and different question formats to assess how biased 17 different LVLMs are.

This paper presents Label Hierarchy Transition (LHT), a deep learning framework designed to improve hierarchical classification, which is the task of sorting objects into multi-level category structures (like organizing a bird into order, family, and species levels). Unlike existing methods that break hierarchical classification into separate classification tasks, LHT uses a transition network and a confusion loss to better capture the relationships between categories at different levels of the hierarchy. The researchers tested their approach on benchmark datasets and a skin lesion diagnosis task, showing improvements over existing methods.

AI models like Mythos are making cyberattacks faster and more dangerous by shortening the time between when security flaws are discovered and when attackers exploit them. Security leaders (CISOs, chief information security officers) need to prepare urgently for this new threat environment where attacks happen at high speed.

This article discusses how AI companies like Anthropic use marketing to promote their capabilities, using Claude as an example of technology that may be overhyped despite being genuinely advanced. The piece cautions readers against getting swept up in marketing claims about AI's power without critical evaluation.

AI is transforming threat detection by processing massive amounts of security data and identifying suspicious patterns faster than humans alone, with 50% of threat detection platforms expected to use agentic AI (AI systems that can take independent actions) by 2028. Organizations are already automating routine tasks like alert review and investigation work, seeing 40-50% efficiency gains for lower-level security operations, while AI agents reduce alert fatigue by clustering similar alerts and prioritizing them based on risk.

AI is moving from experimentation to production deployment in cybersecurity, and security leaders must treat it as a fundamental shift in how security operations work, not just an added tool. Attackers are using AI to conduct faster intrusions (some occurring in under 30 seconds), which exceeds the speed of human-only security responses, making AI deployment urgent for defenders. There is currently a limited window where defenders and attackers have roughly equal access to AI technology, but advantage will go to those who operationalize it most effectively and quickly.