AI Cyber Defense / Cloud, Kubernetes and software supply chain
AI Cyber Defense: Detect permission expansion before data access
Require review for high-impact permission changes and alert on unplanned expansion followed by sensitive access. Retain previous policy versions for a controlled correction.
Environment and inputs: Cloud control planes, workload identity, Kubernetes clusters, CI/CD pipelines and immutable deployment artifacts; provider support varies.
Scenario and goal
An attacker may modify a role or policy before accessing resources that the original identity could not reach.
Why it matters
For this scenario, an attacker can automate permitted API operations after gaining a workload or pipeline identity. Review who can change trust, who can assume the resulting identity and what resources it can reach. Keep source review, build provenance, deployment authorization and runtime permissions as distinct checks. A configuration object or successful deployment does not prove the intended runtime boundary is enforced.
Evidence to collect
Join policy versions, role membership changes, administrator sessions and subsequent resource events using immutable principal identifiers.
Validate with an authorized exercise
Replay a synthetic policy change followed by an export and a legitimate approved change; verify the unplanned chain receives escalation with the exact permission difference.
Rollout and checks
Evaluate effective permissions and runtime behavior in a disposable account or namespace before production rollout. Retain the approved source revision, resulting artifact and independent audit trail. Plan recovery if identity, policy or admission dependencies fail. Reconcile running assets after the change because resources created outside the normal pipeline can otherwise remain outside its checks.
Scope and limits
This guide proposes a defensive test; it is not a tested production design. Check platform support and local requirements. References explain the mechanisms but do not certify the design. Behavior alone does not prove AI involvement.