“ai-safety”
Human-on-the-Loop vs Human-in-the-Loop: Understanding the New Age of AI Oversight
A deep dive into Human-in-the-Loop and Human-on-the-Loop AI oversight models, exploring when each approach is appropriate and how they shape the future of human-AI collaboration.
Sanitizing Agent Inputs: Prompt Validation Before Inference
Protect PydanticAI agents against prompt injection, jailbreaks, and malformed inputs using typed input contracts and pre-inference sanitization filters.
Guardrails for Agent Input and Output: From Validation to Policy
Screen input, validate output, and police tool calls with guardrails. Learn input, output, and tool guardrails in the OpenAI Agents SDK—and the cost of parallel vs blocking execution.
The "Common-Sense" Benchmark: Testing Gemini 3 in Highly Ambiguous Scenarios
Bees drop, wet floors, no chairs left. Build a Common-Sense Benchmark that tests Gemini 3 where the answer is ambiguous, physical, and context-dependent — and where confident is dangerous.
The "Ethics-by-Design" Framework: Embedding Values into Gemini 3 Agents
Don't bolt ethics on after the fact. Apply an Ethics-by-Design framework that embeds values, safety constraints, and audit trails directly into the Gemini 3 agent lifecycle.