Insights
Our insights on AI agent and robotics safety, AI security research, development, and implementation.
Articles
OpenAI's AI Standards Proposal: Four Questions for Companies
Four questions for AI deployment decisions, prompted by OpenAI's standards proposal. What evidence to keep for evaluations, approvals and incident reporting.When AI Development Tools Become the Entry Point: Settings and Permissions to Review
Google's threat intelligence team reported attacks on AI coding assistants and MCP servers. What to check across external tools, credentials, and execution permissions in a development environment.Threat Analysis for AI Systems with Image Inputs
Adversarial inputs to image-recognition AI and image-borne prompt injection against multimodal AI: how they differ from ordinary variation, how to build a threat model, and what to measure.Security Design for Physical AI
Safety and security for Physical AI: separating proposal, authorization, and execution, protecting ROS 2 communication, validating commands, and deciding fault responses at the design stage.Audit Log Design for Robotics Systems
What a robotics audit log should record: how it differs from ROS 2 debug logs and rosbag, the fields linking decision, authorization, and execution, time and ordering, tamper detection, retention.Assessing Person-Detector Misses: An AI Security Perspective
For teams deploying or improving person-detection systems. Our study measures public detectors' detection rates on the same scenes under conditions that include textures built to suppress detection. We show how its method supports an AI security evaluation that grounds threat scoping, verification priorities, and countermeasure decisions in evidence.