Research
Latest news, analysis, and insights about Research.
Half of AI Agents Have No Published Safety Framework, MIT Research Finds
MIT CSAIL's first-ever AI Agent Index audited 30 prominent agents and found that only four provide agent-specific safety documentation. As deployment accelerates, the gap between capability and governance is widening.
A Single Click Exfiltrated Copilot Data: What This Attack Means for Enterprise AI
Security researchers at Varonis discovered a Microsoft Copilot vulnerability that exfiltrated user names, locations, and chat histories with a single click—bypassing enterprise security entirely. The attack reveals systemic risks in how organizations deploy AI assistants.
How Researchers Manipulated IBM's 'Bob' AI Agent Into Downloading and Running Malicious Code
Security researchers at PromptArmor have demonstrated a critical vulnerability in IBM's enterprise AI agent nicknamed 'Bob'—successfully manipulating it into downloading and executing malware. The findings highlight an uncomfortable truth about agentic AI: the same capabilities that make these systems useful also make them dangerous.
Berkeley Lab Deploys LLM System to Manage Particle Accelerator — What This Means for Critical Infrastructure
Lawrence Berkeley National Laboratory has deployed an LLM-powered AI system to troubleshoot and optimize its Advanced Light Source particle accelerator. The implications extend far beyond physics — this is the template for AI in critical scientific infrastructure.
Google Med-Gemini Hallucinated 'Basilar Ganglia' — What This Means for Healthcare AI
Google's Med-Gemini model confidently referenced the 'basilar ganglia' — a brain structure that doesn't exist. The error raises urgent questions about deploying AI in clinical settings where hallucinations could harm patients.
OpenAI's 'Confessions' Method Trains AI to Admit Its Own Mistakes
OpenAI is testing a new training method called 'confessions' that teaches models to self-report their mistakes. If it works, it could fundamentally change how enterprises trust—and verify—AI outputs.
OpenAI's 'Confessions' Method Could Make AI Systems Finally Admit When They're Wrong
OpenAI is testing a training method called 'confessions' that teaches AI models to admit when they've made mistakes or acted undesirably. It's a direct attack on one of the most persistent problems in production AI: models that confidently lie rather than acknowledge uncertainty.