Minimum Viable AI Security: 9 Steps to Build Your AI Security Program in 90 Days

CEOs want every AI tool adopted, teams have experimental AI budgets, and agents are entering through browsers, IDEs, SaaS apps, and cloud platforms all at once.

For CISOs, the challenge is enabling this adoption while maintaining visibility and control, often with limited budgets and resources.

The answer is not to solve every AI security problem at once. It is to establish the right controls in the right order.

At Aktonomy Fall '26, the biggest Agentic AI security summit organized by Akto, security leaders from insurance, healthcare, and SaaS agreed on the goal for the first 90 days: establish visibility, prioritize risk, enforce meaningful controls, and improve continuously.

The Three Non-Negotiables of AI Security

If you're a CISO with 90 days and a limited budget to secure AI, where should you start?

Regardless of how your organization adopts AI, three controls should form the foundation of your security program:

1. Visibility: Know what's running.

Discover the AI agents, tools, and applications operating across your organization, including shadow AI, employee AI usage, and internally developed systems.

2. Data Access: Control what AI can access.

Understand what sensitive data AI systems can access, process, and expose. Enforce appropriate permissions and prevent unauthorized data disclosure.

3. Actions: Govern what AI can do.

Define what AI agents are authorized to access, invoke, and execute. Use runtime controls to detect and prevent unauthorized or risky actions, with human approval for sensitive operations where necessary.

These three controls provide the foundation.

Your 90-Day AI Security Roadmap: 9 Steps Across 3 Phases

Phase 1: Discover and prioritize (days 1 to 30)

1. Assess Your Existing Security Capabilities

Build on existing investments in identity, endpoint protection, DLP, cloud security, and monitoring while identifying gaps specific to AI.

• Assess whether current tools cover employee AI usage, AI agents, MCP servers, and GenAI applications.

• Identify gaps in AI discovery, runtime enforcement, and security testing.

• Prioritize additional capabilities based on business risk.

Outcome: A capability-gap assessment that identifies where existing controls help and where additional AI security capabilities are needed.

2. Establish Visibility Across Your AI Attack Surface

You cannot effectively manage risks you cannot see. AI usage extends beyond approved enterprise tools to browser assistants, desktop applications, developer tools, and locally running agents.

• Discover AI tools, applications, agents, and MCP servers across the organization.

• Identify shadow AI and unapproved integrations.

• Document ownership, business purpose, data access, and approval status.

Outcome: A maintained inventory of employee AI usage and internally developed AI systems.

3. Prioritize Risks Based on Business Context

AI security priorities vary by organization. Companies focused on employee productivity may need to prioritize shadow AI and data exposure, while AI product companies may need to focus on customer-facing applications and autonomous agents.

• Classify use cases by data sensitivity, permissions, autonomy, and business impact.

• Prioritize agents that can access production systems or execute consequential actions.

• Create a remediation plan for the highest-risk use cases.

Outcome: A risk-ranked list of AI systems with clear remediation priorities.

Phase 2: Establish Core Controls (Days 31–60)

4. Define Approved AI Tools and Acceptable Use Policies

Employees need clear guidance on which tools they can use and what information they can share.

• Define approved, restricted, and prohibited AI tools.

• Establish rules for handling sensitive data.

• Create a simple process for reviewing new tools and use cases.

• Communicate policies and enforce restrictions where necessary.

Outcome: An AI acceptable-use policy and an approval process employees can follow.

5. Enforce Least Privilege Across AI Agents

An agent's risk depends not only on what it says but also on what it can access and execute.

• Review agent credentials, API keys, tokens, and connected tools.

• Assign accountable owners and restrict permissions to what each agent needs.

• Review access to MCP servers, APIs, databases, and internal systems.

• Require human approval for sensitive or high-impact actions where appropriate.

Outcome: Documented ownership, scoped permissions, and appropriate action boundaries for high-risk agents.

6. Protect Sensitive Data Across AI Interactions

Sensitive information can be exposed through prompts, model responses, connected data sources, and tool calls.

• Define what data can be shared with external AI services.

• Review providers' commitments on data collection, training, and retention.

• Apply appropriate DLP guardrails and data access controls.

• Monitor for sensitive data exposure across AI interactions.

Outcome: A data protection baseline covering access, sharing, monitoring, and response.

Phase 3: Test and Operationalize (Days 61–90)

7. Test AI Systems Against Realistic Attacks

Traditional security testing remains essential, but AI applications introduce risks such as prompt injection, unintended tool usage, and sensitive data disclosure.

• Threat-model high-risk AI applications and agents.

• Conduct AI red teaming to uncover vulnerabilities such as prompt injection, unauthorized actions, and data leakage.

• Use OWASP guidance to structure AI security assessments.

• Prioritize remediation based on potential impact.

Prompt filtering and jailbreak detection can help, but they are not substitutes for secure architecture and appropriate permissions.

Outcome: A risk-based testing plan with documented findings, remediation owners, and retesting criteria.

8. Establish Ownership and Governance

AI security needs clear accountability and repeatable processes.

• Assign business and security owners to high-risk AI use cases.

• Define approval, review, and incident-response responsibilities.

• Use frameworks such as the NIST AI Risk Management Framework, ISO/IEC 42001, and OWASP guidance where relevant.

• Align AI governance with business needs and regulatory requirements.

Outcome: A lightweight governance model with owners, risk classifications, and review checkpoints.

9. Build a Continuous Improvement Loop

AI systems evolve quickly, so security controls must evolve with them.

• Track inventory coverage, high-risk findings, remediation progress, and permission reviews.

• Reassess systems when models, integrations, or capabilities change.

• Review incidents and near misses to identify control gaps.

• Upskill security teams through hands-on experimentation and targeted training.

Outcome: An ongoing improvement roadmap with measurable priorities for the next quarter.

Build the Foundation for AI Security

Your first 90 days will not eliminate every AI risk. They should establish the foundations of a functioning program: visibility into AI usage, clear ownership, risk-based prioritization, appropriate access controls, data protection, and ongoing testing.

As CISOs emphasized during Aktonomy Fall '26, priorities depend on how an organization uses AI. The objective is to understand your exposure, address the risks that matter most, and invest in additional capabilities where they are needed.

A minimum viable AI security program is not about doing less. It is about doing the right things first, then building on that foundation as AI adoption grows.