Beyond RBAC
Roles answer who. Agents need why.
Why AI agents need access control that evaluates purpose, context and behaviour.
Kevin Keller · kevinkeller.org
Why it matters
RBAC can't see what an agent is doing with its access
- For a human, purpose can be inferred from job function.
- At step 47 of an autonomous chain, that inference breaks.
- The access system knows the role, not the purpose, chain depth or agent type.
Can this agent invoke this tool on this data for this purpose?
A concrete example
Every tool call passes RBAC while the data walks out
- A support agent is asked to summarise a customer's tickets.
- A prompt injection in one ticket triggers
export_customer_data. - Then
call_external_apisends the data out.
Same role, same permissions. The purpose changed and nothing noticed.
The idea
Evaluate context, not just identity
Attribute-based access control asks what the agent is doing, why, with what, and whether that makes sense right now.
Four attribute dimensions
A policy needs far more than a role
- Agent identity: human principal, agent type, version, current tool
- Request context: purpose, data classification, chain depth, prompt lineage
- Environment: time, region, network path, regulatory regime
- Behaviour: access pattern, velocity, data volume, escalation
Three decisions
The policy engine decides more than allow or deny
| Decision | Question |
|---|---|
| Tool authorisation | Can this agent call this tool, on this data, for this purpose, at this depth? |
| Inference routing | Can this data go to this model? |
| Response transformation | Should the result be masked, redacted or sampled for this purpose? |
State of the industry
The pieces are shipping, but in silos
| Layer | Status (March 2026) |
|---|---|
| Agent identity and delegation | Production |
| Tool authorisation | Production |
| Data access with agents as identities | Production |
| Purpose and intent classification | Emerging |
| Behavioural baselines | Early |
| Unified policy engine | Emerging |
What's missing
The connective tissue between layers
- No unified evaluation: no single engine sees tool, data, inference and behaviour at once.
- No shared attribute schema:
agent_type,chain_depthandpurposemean different things per vendor. - Enterprises map between incompatible models at every integration point.
The enterprise pattern
Centralised policy, federated enforcement
| Layer | Owner |
|---|---|
| Global policies, non-overridable | Centre of Excellence: compliance, legal, CISO |
| Domain policies, extend but never weaken | Domain owners |
| Enforcement point configuration | Platform teams |
| Audit and response across all points | Security operations |
Why now
Regulators ask what the AI does with its access
- EU AI Act Article 9 expects governance beyond RBAC for high-risk systems.
- DORA requires automated systems to operate within defined bounds.
- "The agent has a role that can access the table" satisfies neither.
The takeaway
Three attributes, one deny decision, zero data leaked.
Kevin Keller · kevinkeller.org