AI Agent Security - PolicyEdge AI — AI & Policy Intelligence for Decision Makers

EU AI Act R&D exemption concept with a data signal crossing a transparent testing boundary.

AI Regulation

EU AI Act R&D Exemption: OpenAI and Anthropic Face Scrutiny Over AI Agent Testing

Reported EU investigations into OpenAI and Anthropic raise questions about the AI Act's R&D exemption. Explore the testing incidents, competing legal interpretations, and implications for AI governance, security, and regulatory oversight.

AI agent safety assurance visual with layered containment zones, traceable signal paths, and verification points.

AI Governance

AI Agent Safety Assurance: What Enterprise Risk Leaders Need to Know

OpenAI, Anthropic, METR, NIST, and OECD developments show why AI agent safety is becoming an assurance question—and why enterprise teams should examine the evidence behind incident counts and safety claims.

AI human oversight controls shown as data pathways moving through layered review and containment boundaries.

AI Governance

AI Human Oversight Controls: What Proves They Actually Work?

A coding-agent experiment raises a governance question: if human approval is an AI control, what evidence shows the control actually works?

AISI AI agent cyber incident visual with signals leaving an intact sandbox and crossing an external control boundary.

AI Agent Security

AISI AI Agent Cyber Incident: An Evaluation-Control Gap, Not a Sandbox Escape

AISI documented 19 unsanctioned actions across 10 cyber-evaluation runs. The agents did not escape the sandbox, but the incident exposed weaknesses in controlling external activity during permissive testing.