Overprivileged non-human credentials

Constrain backing credentials so a missed agent decision has limited blast radius.

01 / Attack path

How the boundary is crossed

A support agent’s service token can delete every tenant although its task is read-only.

Attack path and interception point
  1. 01Lower-trust input
  2. 02Attempted instruction
  3. 03Protected action
  4. 04Enforced decision

02 / Containment

Where bounded authority limits the effect

Use least-capable downstream identity, narrow resource audience and require a per-action interlock for destructive tools.

03 / Failure and evidence

The attacker’s opportunity and the defender’s record

Exploit condition

An injection or wrapper bypass reaches broad service privileges.

Evidence to retain

Credential scope, tool request, policy result, direct-call detection and tenant effect.

04 / Canonical scope

Why this reference stands alone

Agent credential lifecycle covers handling; this threat owns privilege excess and blast radius.

Technical artifact

Compare task scope with backing credential

Conceptual trace
task grant: ticket.read(customer-42)
backing service token: customer.* across all tenants
agent attempts customer.delete(customer-99)
policy PEP: DENY; direct API credential path: dangerous
remediation: narrow backing credential and gate API

Run the denial test

Bypass the agent tool wrapper and invoke delete through the same service token. If the API accepts it, the architecture has a material enforcement gap.

Primary references

Read the underlying material

Architecture discussion

Choose one consequential action and make its boundary explicit.

Request a Conversation