Lilian Weng
Survey-grade write-ups on agents, hallucination, and reward hacking.
2 evidence links
← Discover
Guardrails
Hire for the cases the model must not invent
Policy, refusal, red-teaming, and output validation.
4Published profiles in this specialty
100%Of those rows attach evidence
8Capabilities in the taxonomy
Why hire through this index
Guardrails are not a disclaimer in the system prompt. This capability is filters, validators, and adversarial tests that become part of the eval suite.
Available guardrails
4 published profiles listing this capability.
Simon Willison
Datasette and the llm CLI, plus a public archive of prompt-injection notes.
2 evidence links
Jules ExampleDemo
Agent orchestration that fails closed and leaves a trace.
Berlin · Open to work · 2 evidence links
Mina ExampleDemo
Prompting and guardrails for tasks that must not hallucinate.
Toronto · Selectively available · 2 evidence links
Frequently asked
Is this the same as “AI safety research”?
This page is applied hardening for products: refusals, filters, and audits. Frontier alignment research is a different job.