Identity · CA
CSA's Fix for Standing Agent Privileges: Make Every Access Request Prove Itself
The Cloud Security Alliance's new Agent Identity Governance Framework doesn't just catalog the accountability problem — it proposes a specific architectural fix: kill standing privileges for agents entirely.
The framework's centerpiece replaces the model where an agent is provisioned once and keeps broad access indefinitely. Instead, it specifies a just-in-time model that replaces standing agent privileges with intent-declared, time-bound, scope-limited grants, fundamentally changing the relationship between agent execution and privilege. Under this approach, an agent doesn't hold a credential that lets it do anything within its role forever. It has to declare what it intends to do, get a scoped grant for that specific task, and have that grant expire.
The framework builds this on top of existing identity standards rather than proposing something new from scratch. It evaluates how OAuth 2.0 and OpenID Connect, SPIFFE/SPIRE, and SCIM can be applied to agent identity governance, identifies the gaps these standards exhibit in agentic contexts, and specifies extensions or compensating controls required to close those gaps. None of those protocols were built with autonomous, delegation-chaining, sub-agent-spawning systems in mind, so the framework treats them as a starting substrate rather than a finished answer.
The numbers motivating this are stark. Only 28 percent of organizations can trace AI agent actions back to a human sponsor across all environments. Fifty-one percent report no clear ownership or accountability for their AI and non-human identity populations, which means nobody is positioned to revoke privileges when an agent's purpose shifts or notice when it drifts outside its original scope. The accountability chain that makes human IAM governable — every account tied to a person who can be contacted or fired — simply doesn't exist for most production agents.
What makes the just-in-time proposal notable isn't novelty — scoped, ephemeral credentials are an old zero-trust idea — but the explicit claim that standing privilege is the wrong default for agents specifically, not just a hardening measure applied on top of it. If an agent has to declare intent before it gets access, every grant becomes a record of what the agent said it was going to do, which is a different kind of accountability than after-the-fact log review. Whether vendors build to this is a separate question from whether the framework is right that it's necessary.
The harder parts — ownership assignment, cross-environment identity propagation, what happens when an agent spawns sub-agents mid-task — are gaps the framework names but doesn't fully solve. It's a design target, not a shipped standard.