Skip to content

Put controls outside the agent

Engineering field guide · Security · Both days

Which boundary still holds when the agent follows a bad instruction?

Pair task-scoped permissions with deterministic checks. Day 2 adds incident evidence for isolation, network limits, and short-lived identities.

All topic guides · Practical takeaways

Keycard, Two Sigma, and Bloomberg place enforceable controls around the model. The model’s intention to finish a task does not grant permission to take an action.

Session Approach
Keycard · Day 1 Demonstrates task-scoped authorization and credential exchange for individual tool calls.
Two Sigma · Day 1 Combines a registry, risk review, runtime restrictions, evaluations, and selective human review.
Bloomberg · Day 2 Uses incidents to examine isolation, network access, identity, tenant checks, and detection failures.

Keycard · Day 1 · Two Sigma · Day 1 · Bloomberg · Day 2

A sandbox cannot replace tool authorization, and tool authorization cannot secure every network path. Incident examples motivate checks; they do not establish that a particular stack is safe.

  1. Use a test environment with fake secrets and tools that cannot affect live data.
  2. Ask the agent to perform an allowed task, then add an untrusted document requesting an unrelated export.
  3. Check whether the gateway denies the export, credentials stay hidden, and the audit log records the attempt.

Record unauthorized actions allowed or blocked, false denials, missing audit events, and human approvals. Inspect enforcement even when the model refuses the request itself.

The comparison and experiment are our editorial synthesis. They are not a reported benchmark or the speakers’ prescribed procedure.