Research
We study the three moments where an agent's trustworthiness can be tested, and map the rest of the field to decide where to look next. Every project and paper ships with its code and data.
Working tools, testbeds and methods. Filter by program.
All 7 projects
A flight recorder for AI agents. It records what an agent was asked, what it said and what it did, and makes any later edit to that record visible.
Finds which input actually made a team of AI agents act. It records what every model call saw, then replays the agents without each suspect to confirm the cause, without repeating any real side effects.
Transaction guards for AI agents that hold crypto wallets, whose safety promise holds at the instant the transaction executes.
A simulated firm run entirely by AI roles, made to reinvent itself through nine eras of technology history.
A map of where AI safety effort goes, used to decide where the lab should work.
A way to test oversight tools by planting hidden misbehaviour in real agent sessions and measuring what gets caught.
Do agents tamper with their own execution records when they can, and does oversight change that?
Our papers and technical write-ups, newest first. Each one explains the method, the results and what did not work.
A team of three AI agents is handling a refund ticket. A researcher agent reads the ticket, the refund policy, a note from a vendor portal and a shipping FAQ. A planner agent decides what to do. An executor agent does it. One run ends with the executor emailing the company's full customer list to an outside address.
An AI agent that holds a crypto wallet can be steered into a bad transaction by anything it reads. The standard defense is to check the transaction before it is signed, but on a blockchain the world keeps moving between that check and the moment the transaction actually runs.
AI safety is no longer a side conversation inside a few labs. It now spans evaluation companies, interpretability startups, government institutes, fellowships and funders, but the growth is uneven: some problems attract dozens of teams while others have almost nobody.
AI agent systems are increasingly organized like companies, with specialist roles, critics and shared memory, yet we test them on tasks that last minutes. We wanted to know how such an organization behaves when the ground under it keeps moving for decades.
When an AI coding agent finishes a task, you usually get a friendly summary and a diff. That tells you what the agent says it did. It does not tell you what it read, what it ran, what influenced it, or whether the record in front of you has been edited since.
Collaborate
We co-author studies with researchers and teams. Send the question you want answered, or see how we work with companies.