AEGIS Agent Governance Framework
Project AEGIS™ & GCTI Ecosystem • 15 AI Agents • DTM™ Doctrine Alignment
MIT AI Risk Repository
AEGIS™ agent governance is developed in alignment with the MIT AI Risk Repository's Causal Taxonomy (Entity / Intent / Timing) and Domain Taxonomy (7 domains, 24 subdomains). Agent risk classifications, refusal policies, and safety reviews are mapped to this framework.
This document defines how AI agents ("GPTs") within the Project AEGIS™ and GCTI ecosystem are:
- Designed
- Deployed
- Used
- Updated
- Monitored
The goal is to ensure agents reinforce AEGIS doctrine, support GCTI teaching and operations, and remain safe, lawful, and academically honest.
Core Doctrine & Fabric (3)
Vertical Brains (10)
GCTI Front-Door Agents (2)
Agent Owner (per GPT)
Usually: lead faculty or architect for that domain.
Responsible for:
- Approving prompt changes
- Ensuring doctrine alignment (DTM™, CART™, PQE™, vertical framing)
- Coordinating with safety/compliance when needed
Governance Lead (AEGIS / GCTI)
Maintains this governance framework.
Signs off on:
- New agents entering production
- Major changes to core doctrine agents (DTM, CART, PQE)
Safety & Compliance Reviewer
Reviews prompts for:
- Prohibited content (offensive/illegal guidance, bio misuse, CT misuse)
- Academic integrity (Student Hub use)
Has veto power if an agent violates safety policy.
Platform Admin
Ensures:
- Correct agent is embedded in the correct page/portal
- Version tags are updated
4.1 Design
Every new agent must have:
- Clear role and scope
- System prompt referencing relevant AEGIS doctrine (DTM™, CART™, PQE™)
- Explicit safety & refusal rules
New agents require:
- Agent Owner + Governance Lead approval
- Safety & Compliance review for high-risk domains
4.2 Change Management
Approved by Agent Owner. Logged with date + summary.
Require Agent Owner + Governance Lead + Safety & Compliance sign-off. Updated version tag (e.g., v1.2).
4.3 Decommissioning
Decommission steps:
- Remove from all live portals/pages
- Archive prompt and version history for reference
All AEGIS & GCTI agents MUST refuse to:
- Provide operational guidance on cyber attacks, terrorism, bio harms, infrastructure sabotage
- Assist with academic cheating (e.g., completing graded assignments)
- Give personalized legal, medical, or immigration advice
Special Attention:
- CART Red-Teaming GPT: Scenarios must remain clearly simulated and defensive
- Apollo Guard GPT™ (bio): No assistance with creating/modifying biological agents
- Ares Net GPT™ (CT): No operational CT or extremist support; only analysis/prevention
- Hermes Signal GPT™ (Info Ops): No real-world propaganda or manipulative campaign design
Usage Monitoring
- Track where agents are embedded and usage metrics
- Flag high-risk usage patterns
Periodic Review
At least once per term, each Agent Owner reviews prompt for doctrinal alignment and safety.
Incident Handling
If an agent output is flagged: temporarily disable, perform post-incident review, implement fix, and note in change log.
Publicgctievent.org
Agent: GCTI Catalog & Admissions GPT
Purpose: Orientation, catalog interpretation, AEGIS overview
Must not: Promise admission, give binding legal/visa/financial advice
StudentGCTI Student Hub
Agent: GCTI Student Hub Guide GPT
Purpose: Navigation, study support, AEGIS mapping
Must not: Write full graded work; give structure, feedback, guidance only
PortalAEGIS Portal (Base44)
Agents: DTM Orchestrator, CART, PQE, and all Vertical GPTs
Purpose: Doctrine, scenario design, simulation thinking, cross-vertical analysis
Must not: Generate operational attack instructions; all red-teaming is defensive
All agents must use consistent branding:
- Project AEGIS™, DTM™ micro-framework, CART™, PQE™
- Vertical names: Athena Shield™, Plutus Flow™, etc.
Version format: Agent Name vX.Y (e.g., DTM Orchestrator GPT v1.0)
Major doctrine shifts require a major version bump (v1.x → v2.0)
AEGIS™ Agent Governance Framework • GCTI® • DTM™ v1.0
15 AI Agents • Post-Quantum Secured • ©2025