How to find agent AI risks with black-box red teaming
Beyond chat evals, a taxonomy-driven framework that automatically finds multi-step risks in agentic AI from the outside
Blog
Stories of AI technology, research, products, and the people shaping them
Thoughtful coverage of technology and product design and validation, notable papers and AI trends, Patty’s new initiatives, and the culture behind the work.
Latest
Page 3
The menu is an execution prior, state path tool menus for online agents
Why ranking tools by similarity to the request leaves the final action stranded, and how a menu built as a prior over execution paths lifted online success on ToolBench.
Not saving experiments but saving decisions
A scale-up reading of PNNL CICERO records through active learning. The gap between 16 wells and 48 wells on recycled NdFeB, the SmCo purity yield tradeoff, and choosing batches by Bayes risk.
What changes when AI learns entangled instead of smart alone
Walking through a paper that models entangled collective behavior as probability waves and tests it on Chinese intraday trading data, and asking why AGI might need entangled game modules, and where the limits sit.
Choosing a Less Harmful Route Through the Arctic, Not Just the Fastest One
A walkthrough of a human in the loop, multi-agent GeoAI framework that treats Arctic routing as a multi-criteria choice across operational, physical, ecological, and community concerns.
Subagents and agent skills, two ways to run knowledge that has to last
A reading of when in-context skill execution frays and when subagent delegation holds, through contracts, procedure, and coordination cost.
Rewriting the statistical mechanics of learning in function space
Reading a theory that treats functions and their operators as macroscopic variables for learning dynamics, and tracing why error and fluctuations settle along smooth directions.
RobustSGPO: How to Govern Search Space When Agent Harnesses Evolve
A walkthrough of RobustSGPO, which repairs the open choices in SGPO with explicit edit requests, patch construction and checks, and restarts from retained snapshots, and why that lifted completion and quality in the AgentX brainstorming flow.
When the scores match but the work does not
Behind success rates in the eighties, 558 trajectories reveal different habits of reasoning and tool use