Guarding Delegated Authority: Professional certification and operational safeguards for AI agents

AI agents can perform consequential actions through tools, credentials and delegated permissions. The governance problem is not only whether a threat can be recognised, but whether authorised limits remain enforceable while evidence is incomplete and intervention is time-sensitive. This discussion paper proposes two specialist qualifications within one professional scheme: an AI Red-Team Assessor, responsible for authorised adversarial evaluation, and an AI Security Operator, responsible for operational monitoring, containment and recovery. The proposal extends the author's Clean Engagement, Reflection Skills, Verification and Provenance and Certified Remote Operator frameworks. It distinguishes individual competence from service capability and deployment assurance, and argues that decisive restrictions must be enforced outside the model. A staged pilot is proposed for enterprise agents using documents, email and cloud business tools, with independent assessment and measures of containment, false intervention, evidence quality, recovery and delivery cost. The paper is a conceptual design proposal, not an empirical evaluation. It establishes neither accreditation, effectiveness, market demand nor commercial returns. Its contribution is a testable connection between professional qualification, delegated authority and operational evidence. Final discussion paper, version 1.0. Conceptual proposal, not an empirical evaluation. Not peer reviewed. No accreditation, endorsement, measured effectiveness, market demand or commercial returns are claimed. The sole author, B. Fischer, leads the GCTI/Emboa initiatives discussed and has a commercial interest in their development. AI assistance was used in drafting, editing, source checking and document preparation; it is not independent assurance.