AI Sec Watch

Microsoft and OpenAI have removed a clause from their partnership agreement that previously governed what would happen if AGI (artificial general intelligence, an AI system that can do any intellectual task a human can do) was developed. Under the new terms, Microsoft remains OpenAI's primary cloud partner with first access to new products, but OpenAI now has freedom to use other cloud providers instead of being locked into Microsoft's Azure platform.

Elon Musk, a cofounder of OpenAI, is suing the company and its leaders Sam Altman and Greg Brockman, claiming they abandoned OpenAI's original mission to develop AI for humanity's benefit and shifted focus to profit instead. OpenAI counters that the lawsuit is a baseless attempt by Musk to harm a competitor to his own AI ventures. Musk is seeking the removal of Altman and Brockman, an end to OpenAI's nonprofit status, and up to $150 billion in damages.

OpenAI has received FedRAMP 20x Moderate authorization (a security certification that allows U.S. government agencies to use cloud services), making ChatGPT Enterprise and the API Platform available for federal use. This certification was achieved through a faster authorization process that emphasizes cloud-native security evidence and automated validation, allowing government agencies to access advanced AI capabilities like GPT-5.5 while meeting federal security and governance requirements.

Qualcomm is reportedly partnering with OpenAI and MediaTek to develop custom smartphone chips, with mass production expected in 2028. According to analyst Ming-Chi Kuo, OpenAI believes controlling both the operating system (the software that runs a device) and hardware will let it deliver comprehensive AI agent services (AI systems that can perform tasks autonomously) that use real-time smartphone data to improve performance.

LVMark is a watermarking technique that hides invisible digital markers in video diffusion models (AI systems that generate realistic videos from text descriptions) to protect against unauthorized use. The method improves on existing approaches by maintaining temporal consistency (keeping the watermark stable across consecutive video frames) and preserving video quality through a specialized decoder that combines frequency analysis with color information. The technique successfully embeds 512-bit capacity watermarks that remain detectable even when videos are distorted or attacked.

ShadowCoT is a backdoor attack (a hidden vulnerability inserted into an AI model that causes it to misbehave when triggered) that targets Chain-of-Thought reasoning, which is a technique where LLMs show their step-by-step thinking to solve complex problems. Unlike simpler attacks, ShadowCoT hijacks the model's internal reasoning process by subtly rewiring how attention flows through the model and changing intermediate representations (internal data the model creates while processing), allowing it to produce logical-sounding but harmful outputs while avoiding detection.

Sanitizer is a new framework for federated learning (a training method where multiple computers learn together without sharing raw data) that addresses two major challenges: privacy and Byzantine attacks (when some computers send bad or malicious data). The framework uses more efficient cryptographic techniques and two new protocols for combining data from multiple sources, achieving 19-23 times faster performance than existing methods while maintaining accuracy and security.

This research presents HEFLGuard, a security system that detects backdoor attacks (hidden malicious behavior injected into AI models) in homomorphic encryption-based federated learning (a privacy-protecting technique where multiple parties train an AI model together using encrypted data). HEFLGuard works by having the server create validation models from encrypted client groups and having clients compare their local model outputs on trusted samples to identify suspicious behavior, while also using Byzantine fault-tolerant aggregation (a method that functions correctly even when some participants act maliciously) to ensure robustness.

Distributed Machine Learning (DML, where multiple computers train an AI model in parallel) creates challenges in determining which clients contributed what to the final model and preventing dishonest clients from falsely claiming ownership. This paper introduces MAMMON, a system using matrix commitments (a cryptographic method for verifying that data hasn't been tampered with) to verify each client's training history and establish fair ownership of the resulting model while using less computing power than existing methods.