AI Sec Watch

Nvidia is shifting focus toward CPUs (central processing units, the main general-purpose chips in computers) alongside its famous GPUs (graphics processing units) because agentic AI (AI systems that autonomously complete tasks by orchestrating multiple agents working together) requires significant general computing power to move data and coordinate workflows. The company is unveiling new CPU details at its GTC conference, with demand from major partners like Meta driving a predicted doubling of the CPU market from $27 billion in 2025 to $60 billion by 2030.

Anthropic has made 1M context (the ability to process 1 million tokens, which are small units of text that AI models break language into) generally available for its Opus 4.6 and Sonnet 4.6 models at standard pricing, with no additional charge for using the full window. This differs from competitors like OpenAI and Gemini, which charge premium rates when token usage exceeds certain thresholds (200,000 tokens for Gemini 3.1 Pro and 272,000 for GPT-5.4).

ServiceNow's CEO warns that AI agents (software programs that can perform tasks independently) automating work could push college graduate unemployment into the mid-30s within a few years, making it harder for entry-level workers to stand out. Multiple major tech companies are already using AI to cut jobs and reduce hiring costs, affecting both technical roles like coding and white-collar positions across industries.

The US Department of War designated Anthropic as a 'supply chain risk' (a classification that prevents a company from being used in government contracts) after the company refused to remove safety restrictions on its AI model Claude, specifically rejecting military demands to enable fully autonomous weapons and domestic mass surveillance. Anthropic is challenging this designation in court, and legal experts question whether the Department of War has the authority to impose such restrictions outside of actual contract disputes.

Researchers developed SMInject, a new type of attack that tricks multimodal AI models (systems that process both text and images together) by injecting deceptive instructions that exploit how different data types relate to each other. The attack is designed to be harder to detect than previous methods and achieves higher success rates while still keeping the model appearing to work normally.

This research addresses the challenge of detecting deepfakes (synthetic videos or images created by AI to manipulate someone's appearance) by proposing a new detection method called a triple-branch network. The method analyzes images using both spatial features (visual patterns) and frequency features (patterns that emerge when you break down images into their component wavelengths), combined with a mathematical approach based on mutual information theory (a concept measuring how much information one variable reveals about another) to improve detection accuracy across different types of forgeries.

Researchers created Fraud-RLA, an adversarial attack (a method to trick AI systems by finding weaknesses) that uses reinforcement learning (a technique where AI learns by trial and error to maximize rewards) to evade credit card fraud detection systems. The attack is designed to steal maximum amounts of money while requiring less background knowledge about the system than other attack methods. Tests showed the attack worked effectively against realistic fraud detection systems.

Third-party applications (TPAs, external software added to Kubernetes clusters) can be exploited to take over Kubernetes clusters, a container orchestration platform (software that manages containerized applications). Researchers created KubeSec, a tool that automatically analyzes these applications to find security weaknesses, discovering 562 insecure RBAC (role-based access control, a permission system in Kubernetes) patterns and 375 vulnerabilities affecting millions of users. The research revealed that these vulnerabilities take over 10 months on average to fix, highlighting a critical security gap in Kubernetes cluster management.

The US military is considering using generative AI systems (AI models that can create text and analyze data) to help rank military targets and recommend which ones to strike, with human officials making final decisions. The Pentagon is also favoring OpenAI's ChatGPT and xAI's Grok for these high-stakes military applications, while facing criticism from officials who claim that Anthropic's Claude would negatively affect the defense supply chain.