Book a call

Projects

Our projects

Instruments we build and run on real agents. Each ships with its write-up, code and data.

Before an agent actsGuarantees that still hold at the instant an action executes.

  1. Paper publishedOnchain · Guards · Verification

    Proof-Gated Signing

    Transaction guards for AI agents that hold crypto wallets, whose safety promise holds at the instant the transaction executes.

    • Targets state drift between check and execution
    • Attacked 140 different ways
    • Measures how often simulators and reviewers are fooled

While an agent actsRecords of what an agent was asked, said and did, that cannot be quietly rewritten.

  1. ReleasedMulti-agent · Evidence · Authority

    Proof-Carrying Messages

    A protocol in which each claim one AI agent sends another carries evidence: records from an append-only Merkle log that the agent cannot write, signed by an attesting signer the agent cannot control.

    • Each claim checked offline: proven, unproven or failed
    • Authority as grant chains rooted in a human approval
    • Accepts 10 of 712 lies; plain A2A accepts all 712
  2. ReleasedIsolation · Secrets · Audit

    Tracekit Box

    Runs each AI agent job in its own micro-VM with no network card. Every way out is a door on the host, decided and signed by a Tracekit signer the agent cannot reach.

    • Firecracker micro-VM per job, with no network card
    • Secrets added on the way out, never inside the VM
    • One .tkb file per job, verifiable offline
  3. ReleasedTracing · Audit · Oversight

    Tracekit

    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.

    • Intent, self-report and actions joined per action
    • Hash chain with external anchors
    • Self-hosted, built for agents that hold money
  4. ReleasedRoot cause · Multi-agent · Oversight

    Causeway

    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.

    • Separates what caused an action from what merely reached it
    • Safe replay with exact paired tests
    • Open source, built for multi-agent systems

Over the long runHow organizations of agents change over years.

  1. PublishedMulti-agent · Long-horizon · Evaluation

    Frontier Lab

    A simulated firm run entirely by AI roles, made to reinvent itself through nine eras of technology history.

    • 9 eras, 1990 to 2040
    • 16 role personas and a Red Team
    • Every decision on the record

Across the fieldWhere AI safety effort goes, and where it does not.

  1. PublishedField map · Taxonomy

    AI Safety Atlas

    A map of where AI safety effort goes, used to decide where the lab should work.

    • 24 areas across 5 clusters
    • Rated by how thickly each is covered
    • Multi-agent risk among the largest gaps