GenAI COMPASS Playbook · 2025

The
Playbook

Knowing the risks isn't enough - you need a repeatable method to stay ahead as AI keeps changing. COMPASS is 's four-step way of doing exactly that.

Four simple steps: Observe, Orient, Decide, Act. Then do it again.

The Loop

Observe · Orient · Decide · Act

The OODA loop, borrowed from the fighter pilot, applied to the machine. Because in the fast-changing world of generative AI, the side that cycles faster wins.

1 Observe

See the Attack Surface

Build a structured view of your AI-related threats. Map every deployment - cloud, on-premises, hybrid - and every profile: the adversary using AI against you, your own Copilots and Geminis, and your custom agents.

2 Orient

Find Your Bearing

Align what you see with known vulnerabilities, real-world incidents, and threat intelligence. Map each flaw to a CWE and score it. Ask the red-team questions. Understand where you stand.

3 Decide

Choose Your Course

Map threats to defenses and mitigations. Decide what to fix, what to monitor, and what to accept - based on your risk appetite, not your fear.

4 Act

Steer, Then Re-Steer

Turn findings into an actionable, owned, time-boxed roadmap - then loop back and observe again. It is a cycle, not a destination.

The Threat Profiles

The Four Threat Profiles

COMPASS organizes the attack surface into profiles - because the threats you face from the outside are not the same as the ones you breed on the inside.

I
Profile 1 · External Threats

The Adversary Outside

“Attackers using AI against your company.”

+

What it is

Adversarial use of AI against your organization - attack acceleration, identity compromise, and fraud at machine scale.

What to watch for

  • Deepfakes for executive impersonation and vishing
  • AI-powered credential stuffing and password cracking
  • AI-augmented spear-phishing and BEC
  • Invoice forgery and payment redirection
  • Automated disinformation and influence operations

How to fix it

  • Update fraud detection for deepfakes
  • Monitor brand and sentiment
  • Harden authentication (phishing-resistant MFA)
II
Profile 2a · Internal Adoption

The Neglect Within

“Risks from your own team using AI without guardrails.”

+

What it is

Risks from your own AI use - shadow AI, absent governance, and unmanaged model risk.

What to watch for

  • No AI governance board or ownership
  • Shadow AI deployed by business units
  • Overprivileged service accounts and non-human identities
  • No input/output logging for GenAI
  • No rollback plan if the AI fails or is compromised

How to fix it

  • Central model registry and audit trail
  • AI lifecycle metrics and risk ranking
  • Incident playbooks for AI-specific threats
III
Profile 2b · Copilot & Gemini

The Assistant With Too Many Keys

“Your office AI assistant can surface data it shouldn't.”

+

What it is

Enterprise assistants surface data users have access to - but should not. Overprivileged access becomes a search engine for secrets.

What to watch for

  • Overprovisioned access exposed by advanced search
  • Immature or missing data classification
  • Unmanaged AI service accounts
  • Risky default settings, content sprawl

How to fix it

  • Least privilege before enabling the assistant
  • Tiered sensitivity labeling, enforced
  • Govern non-human identities
IV
Profile 2c · Agentic Projects

The Autonomy You Handed Over

“Your own AI agent acting beyond what you meant it to do.”

+

What it is

Agentic and generative projects introduce autonomy, tool misuse, and emergent behavior without oversight.

What to watch for

  • Agents initiating harmful or unauthorized actions
  • Tool misuse via prompt injection or manipulated context
  • No secondary validation of AI outputs
  • One compromised agent disrupting entire workflows
  • No traceability or cryptographic logs of decisions

How to fix it

  • Least agency and least privilege for every tool
  • Human-in-the-loop for irreversible actions
  • Sandbox, segment, and bound blast radius
The Score

Speaking the Industry's Language

To fix an AI vulnerability you must first name it. COMPASS maps each AI weakness to the industry's shared taxonomy - CWE and CVSS - so an LLM flaw is scored and tracked like any other software defect.

Common Weakness IDs (CWE)

  • Prompt injection → CWE-77 Command Injection
  • Jailbreak → CWE-707 Improper Neutralization
  • Code injection → CWE-94
  • Output escaping failure → CWE-116
  • Training-data poisoning → CWE-20 / CWE-494
  • Unvalidated LLM prompting → CWE-1427

How Bad Can It Get?

A zero-shot prompt injection allowing model override can score CVSS 8.6-9.8 (High-Critical) - a remote, low-complexity, no-auth attack with high integrity impact. COMPASS then folds that into a simple 5-point impact/likelihood scale, erring high when unsure.

"The Nuclear AI Disaster" - document the worst-case AI scenario your organization could face. That single scenario becomes the anchor for every control you build.

Do This First

Five Steps to Start Today

1

Confirm legal and regulatory compliance obligations are up to date for AI.

2

Update fraud detection - invoicing, money movement, and hiring - to detect and mitigate deepfakes.

3

Update the incident response plan to include AI incidents, including a disinformation strategy.

4

Review third-party partners and add AI-specific questions to your vendor assessments.

5

Write an AI policy, or update the Acceptable Use Policy, so company data stays out of unapproved tools.

← How Attackers Think How to Fix It →