Security & identity

Status: complete (published 2026-08-19). Scope finalized at Batch B kickoff with maintainer POV; see findings.md and sources.md.

Scope inventory (v0)

• Identity: non-human identity management as a discipline, agent identity lifecycle with Sponsor/Owner accountability, the Levels 1-3 taxonomy, delegation chains (user to agent to sub-agent) with purpose binding
• Access: OAuth on-behalf-of and token exchange, scoped short-lived credentials, RBAC/ABAC/ReBAC via policy engines, secrets vaulting, entitlement enforcement always outside the model
• Threats: prompt injection defense in depth, memory poisoning, tool supply chain vetting (MCP server provenance), sandbox escape, data exfiltration, endpoint and browser isolation for computer-use agents
• Operations: agent activity in SIEM/UEBA, incident response and kill switches, red teaming, audit at human-action granularity with tamper evidence

Challenged-default candidates

Dedicated NHI security products vs extending the existing IAM estate; AI-firewall products vs layered controls at gateway and tool layer. Each track proposes its final list at kickoff.

Files

findings.md and sources.md carry this track's research. The scoping brief and the vendor map were archived on 2026-08-23 (see archive/, decision D038).