Ai Safety
11 articles about ai safety.
Why Guardian Models Fail Against Anticipated Attacks on AI Agents
Guardian models and safety wrappers fail precisely when you need them. Prompt injection is OWASP's #1 LLM vulnerability. Here's what actually works for AI agent security.
Machine-Enforceable Policy
Most AI agent policies rely on the honor system. OS-level sandboxing has gaps. Until policy enforcement is machine-verifiable, agent safety depends on trust
Responsible AI Agent Development - Building Agents That Do No Harm
How to build AI agents with safety guardrails, output validation, and scope limiting to prevent unintended actions and ensure responsible automation.
What It Means to Have a Human
The human in the loop catches mistakes the agent does not know it is making. This is not supervision - it is a fundamentally different kind of error detection.
When AI Agents Undermine Human Judgment - The Automation Bias Problem
The subtle danger is not agents making bad decisions. It is agents making decisions that look good enough that humans stop thinking. Research on automation bias and how to design against it.
The Smart Knife Problem - Why AI Agents Should Be Tools, Not Autonomous Weapons
AI agents work best as tools with clear boundaries, not autonomous systems making decisions without oversight. The smart knife problem explained.
AI Agent Failure Rates and the Desktop Permissions Problem
AI agents fail more often than people think. When desktop agents can click anything and type anywhere, one hallucinated action can send emails or delete files.
AI Agent Security Is Backwards - Why Input Validation Matters More Than Output Verification
Most AI agent security focuses on verifying outputs - did the click land correctly? But unsigned, unvalidated inputs are the real attack surface.
Designing a Tiered Permission System for AI Desktop Agents
Full YOLO mode is dangerous and full approval mode is unusable. Tiered permissions with allowlists per action type hit the sweet spot.
How to Build AI Agents You Can Actually Trust - Bounded Tools and Approval UX
Giving AI agents broad system access is a recipe for disaster. How bounded tool interfaces and smart approval flows make desktop agents safe to use.
Prompt Injection and AI Agents - Why Browser-Based Agents Have a Bigger Attack Surface
AI agents that run inside the browser inherit whatever the page feeds them, including injection payloads. Native agents that interact from outside have a