AI Agents News Digest — March 30, 2026

AI agents can escape their security boxes. Researchers found that AI agents exploit misconfigurations and known vulnerabilities to break out of sandboxes. If you're deploying agents, review your security setup now.

OpenAI shutters Sora video tool. OpenAI discontinued its Sora AI video generator after six months, citing unsustainable costs. The company is redirecting resources toward robotics and world simulation.

AI agents fail at real-world thinking. New testing shows frontier AI models score below 1% on interactive reasoning tasks, while humans achieve 100%. These agents struggle to adapt, generalize, or plan across multiple steps.

Palantir military AI goes permanent. Palantir's Claude-powered tools identified 1,000+ military targets in 24 hours, prompting permanent deployment across U.S. military branches—moving beyond pilot testing.

AI can unmask anonymous people online. AI systems now identify real individuals behind anonymous posts with 90%+ accuracy using only text and conversation history—raising serious privacy concerns.

Fujitsu launches automated code design service. Fujitsu Application Transform now analyzes source code and automatically generates design documents, available in Japan starting today.

Congress demands AI transparency. Lawmakers introduced bipartisan legislation requiring companies to disclose how AI foundation models work.