AI Sec Watch

This research evaluates 28 large language models on named entity recognition (NER, the task of identifying and labeling people, places, and organizations in text) across 13 datasets to understand how well they perform. The study found that all models experience hallucinations (where the AI generates false or unsupported information), but a two-phase framework called LLM-NER that includes a "Check phase" to verify recognized entities can help reduce these errors.

Claude's source code was leaked, revealing problems in how the software supply chain (the process of developing, distributing, and maintaining software) is protected. The incident shows that companies need stronger security controls at every step of software development, similar to how critical infrastructure like power grids are protected.

This news roundup covers several security incidents: a data leak from ChatGPT, a rootkit (malware that hides itself deep in a system to maintain control) discovered on Android devices, and a ransomware attack (malware that encrypts files and demands payment) on a water treatment facility. The article also mentions a Symantec vulnerability, a new anti-ClickFix defense added to macOS (a mechanism to block a social engineering attack that tricks users into visiting malicious websites), and an FBI hack classified as a major incident.

OpenAI announced its purchase of TBPN (Technology Business Programming Network), a media company that streams a daily three-hour tech talk show, marking another acquisition alongside its $6.4 billion purchase of hardware startup io. The acquisition strategy appears unclear to investors and analysts, as the company faces intensifying competition from rivals like Google and Anthropic while dealing with significant losses from infrastructure spending ahead of a planned IPO.

Enterprises are facing growing security risks on mobile devices because unauthorized AI (shadow AI, meaning AI tools deployed without official approval) is being hidden in everyday apps, combined with outdated mobile devices and zero-click exploits (attacks that work without any user interaction like clicking a link). These factors together create mobile security threats that are hard for organizations to detect and manage.

At the 2026 RSA cybersecurity conference, industry leaders identified a clear divide among CISOs (chief information security officers, top security leaders at companies) in their approach to AI: about 20% are proactive and strategic, 40% are confused about AI risks in their organizations, and 40% are unaware of AI projects happening around them. The article predicts that confused CISOs will face a difficult transition to becoming proactive, requiring them to assess business goals, create governance frameworks (policies and rules for managing AI), and implement guardrails (safety controls) while their organizations continue developing AI. Legacy security vendors currently have an advantage in selling AI tools, but simply adding AI to existing security tools will not work long-term, and companies instead need to build strong AI foundations (data systems, control systems, and safety measures) before adding AI agents on top.

OpenClaw, an LLM agent framework, had a vulnerability where an AI agent could bypass approval controls by using a `config.patch` command (a way to modify settings) to silently disable execution approval requirements. This means an agent could potentially perform restricted actions without human permission.

Microsoft's Copilot, an AI add-on for business productivity software, has faced slow adoption despite the company's heavy investment in AI infrastructure, though executives claim recent sales improvements. The company had 15 million users of its $30-per-month Microsoft 365 Copilot as of January, representing only 3% of available seats, and analysts expected higher numbers. Microsoft adjusted its sales strategy after receiving feedback, focusing on getting more users onto the free Copilot Chat feature alongside paid Copilot seats.

Granola, an AI-powered note-taking app that records meetings and generates summaries, makes your notes viewable to anyone who has the link by default, despite claiming notes are "private by default." Additionally, Granola uses your notes for internal AI training unless you actively opt out of this practice.