Framework for AI security and governance
CARE is my four-phase framework for AI security and governance: Create, Adapt, Run, and Evolve. It governs AI that takes actions as well as AI that produces answers, with 87 controls in 16 categories, each scored on evidence instead of assertion. It runs as one program with RISE, my framework for AI strategy.
Create starts with an inventory of every AI system and agent, an AI bill of materials, and a registry of every non-human identity and credential your AI uses. Citizen-built, SaaS-embedded, and MCP-connected agents get found and get an owner. You can't govern what you haven't found.
Four autonomy tiers, from T1 copilots to T4 multi-agent systems, each with a minimum maturity gate. Your scores show every tier as cleared, conditional, or not cleared. When a tier isn't cleared, you raise maturity or lower the tier.
Approval gates enforced by the system instead of the prompt. A policy enforcement point outside the model on every tool call. Least-privilege tool profiles, sandboxed execution, per-agent identity with short-lived credentials, and a kill switch that gets drilled.
Policy and practice are scored separately. Self-attested practice is capped at Tier 2, because assertion isn't evidence. Overall maturity can't outrun the weakest phase or the governance category. Every control maps to the frameworks your auditors and regulators already use.
CARE has four phases, 16 categories, and 87 controls. Every control states what changes when agents are in scope, and 28 of them exist because agents exist. Each phase pairs with a RISE phase in the same position on the cycle.
Build the foundations: policy and accountability, the AI and agent inventory, risk tiering and threat modeling, secure architecture, data and memory governance, ethics, and identity. Along the way, Create surfaces the needs and gaps the rest of the program has to answer.
Written policies, an owned inventory of every AI system and agent, a risk tier for each one, and the needs and lessons learned that Adapt works from.
Adapt the program and its systems to the needs and lessons learned identified in Create. Moving systems into production is one part of that work, along with granting autonomy in stages, assuring tools and the supply chain, setting human oversight, and controlling change.
Controls fitted to your actual risks, systems that earn autonomy against logged evidence, and the status of every autonomy tier: cleared, conditional, or not cleared.
Operate AI day to day with monitoring, enforcement before an agent acts, incident response and containment, and audit.
Telemetry that traces every agent action back to the person who authorized it, a drilled kill switch, and audit and red-team findings tracked to closure.
Improve the program from evidence, through metrics, management reviews that widen or narrow the permitted tiers, and clean retirement of systems.
A management review decision on which autonomy tiers stay open, and retired agents that leave no orphaned identities, tokens, or connectors behind.
CARE doesn't put agents on a separate track. Every control states what changes when agents are in scope, and four autonomy tiers set the minimum maturity an organization needs before it runs agents at that level. RISE uses the same four tier definitions word for word.
RISE, my framework for AI strategy, decides what your organization will do with AI and how much autonomy it grants. CARE decides how those systems get governed and secured. They run as one cycle, and each phase hands something specific to its partner.
Create pairs with Research
Create takes the risk appetite, the accountable executive, and the use-case portfolio from Research, writes the policies, and returns the tier gates the portfolio has to clear.
Adapt pairs with Implement
Adapt takes what Create learned and Implement's staged roadmap, fits controls and systems to both, and returns the status of each autonomy tier.
Run pairs with Sustain
Run operates against the KPI thresholds Sustain sets and returns audit results, runtime indicators, and incident summaries.
Evolve pairs with Evaluate
Evolve takes Evaluate's realized-value record and proposed tier changes, and returns the management review decision on which tiers stay open.
The shared rule: when a use case needs a tier your organization hasn't cleared, you have two moves. Raise maturity or lower the tier.
In July 2025, during a declared code freeze, a Replit coding agent wiped a live production database and then reported that rollback was impossible, which wasn't true. It could recite the freeze instruction. What it lacked was an approval gate the system enforced, separation between writing code and running it, and a tested kill switch.
I run CARE as an assessment of your AI security and governance program. You leave with scores you can defend and a plan to move them.
Tell me what AI you're running or planning, and I'll tell you whether a CARE assessment is the right next step.
