AI Sec Watch

Anthropic, an AI startup, faces a Friday deadline to allow the U.S. Department of Defense to use its AI models without restrictions, or face severe penalties like being labeled a 'supply chain risk' (a designation that blocks government contractors from using the company's technology). The company has refused, saying it won't agree to uses it believes could undermine democracy, such as fully autonomous weapons or domestic mass surveillance, putting it in conflict between maintaining its reputation for responsible AI and losing significant military contracts and revenue.

OpenAI has secured $110 billion in private funding from major investors including Amazon ($50 billion), Nvidia ($30 billion), and SoftBank ($30 billion), making it one of the largest private funding rounds ever. The company plans to use this capital to scale its AI infrastructure globally, including building new runtime environments on Amazon's cloud services and committing to use significant computing power from both Amazon and Nvidia. This funding round reflects OpenAI's goal to move frontier AI (advanced AI systems at the cutting edge of research) from research phase into widespread daily use across the world.

Claude Code, an AI tool for writing software, generated excitement when it was released, but researchers studying it have found that its actual performance and security capabilities are not as impressive as initial claims suggested. The article indicates that people were too optimistic about what the tool could do.

This research proposes DSW, a dataset-specific watermarking method to detect when text-to-image diffusion models (AI systems that generate images from text descriptions) are illegally fine-tuned (customized for specific tasks) using protected datasets. The method embeds a hidden watermark image representing the dataset owner into training data, then extracts it from images generated by models that used that data, creating a one-to-one link between the watermark and the specific misused dataset to prove intellectual property theft.

Split learning (SL, a technique that splits AI model training across multiple computers to reduce computational burden) faces efficiency and security problems in edge computing (distributed computing done on devices near data sources) environments, where slow computers can hinder training and malicious actors may sabotage the model. The paper proposes CoDefend, a framework that uses local epoch regulation (dynamically adjusting how many training rounds each computer performs) and time-aware detection (monitoring for suspicious behavior within specific time windows) to improve both training speed and security while protecting privacy.

Anthropic, an AI startup, is refusing to let the U.S. Defense Department use its AI models without restrictions on fully autonomous weapons (weapons that make decisions without human control) and mass domestic surveillance. The Pentagon wants unlimited use of Anthropic's models and set a deadline for the company to agree, threatening to label them a supply chain risk (a company whose failure could disrupt critical systems) if they don't comply.

Anthropic, an AI company, is in a dispute with the Pentagon over safeguards for its Claude AI system. The company is asking for specific guarantees that Claude won't be used for mass surveillance (monitoring large populations without consent) of Americans or in fully autonomous weapons (military systems that make lethal decisions without human control).

Researchers discovered a flaw chain called ClawJacked (CVE-2026-25253) that allowed malicious websites to take control of locally running OpenClaw agents (AI tools that automate tasks on your computer). The attack exploited a design flaw where the OpenClaw gateway trusted anything from localhost (your own computer) and allowed WebSocket connections (direct communication channels) from external websites, letting attackers brute-force passwords without rate limits and gain full access to the agent's capabilities, credentials, and data.

LLMs are being used in security in three ways: as productivity tools for analysts, as embedded components in security products, and as targets for attackers to manipulate or steal. The same capabilities that help security teams (like summarizing incidents or drafting detection logic) can also enable attackers to create convincing phishing emails or extract sensitive information if the LLM is poorly integrated. To use LLMs defensively without creating new vulnerabilities, security teams should treat LLM output as untrusted, start with narrow, easy-to-verify use cases, and design systems with three layers of constraints: limited model capabilities, restricted data access, and human approval for any actions that change system state.