From Shadow AI to Self-Healing: The DREAM Framework for AI Security

A few months ago, a leading health system asked the following:

“How can we create an environment into which our clinicians can vibe code 1,000 agents this year (2026)? How can we create a self-discovering, self-monitoring, and self-managing ecosystem?”

At first, it seemed like they were dreaming. But this is a reality organizations are working through today. AI offers real promise, but also introduces new risks across our surfaces. Organizations must be brilliant at the basics to mitigate the risks of AI sprawl today and to stay resilient for changes tomorrow. Cybersecurity can be quick to focus on the negative, but with AI it is ok to dream, as long as we cover the basics so we can sleep well at night.

The DREAM framework

The DREAM framework describes an optimal, automation-minded security program built for the evolving AI era. It is intended to be simple, vendor / platform agnostic, and holistic: a flexible rubric to evaluate programs today and prepare for tomorrow.

D: Discover
R: Review Inventories
E: Evaluate Controls
A: Adjust Controls
M: Monitor & Feedback Loop

AI does not care what you were intending to do, what you actually did, or what team you work on. The framework therefore applies across the entire continuum of an organization. It runs from as far left as possible in the SDLC, through the hybrid environments where applications, agents, and employees work, to your edge surfaces, and even outside your organization to your brand exposure.

Discover: Discovery must be automated and holistic across all surfaces, starting with stages of the SDLC. Employees, agents, and external actors should be discovered with identity-based attributes and behavioral baselining. More than 40% of employees are installing AI tools. More than 47% of interactions happen with non-corporate identities, and more than 6% of chatbot conversations contain sensitive data. Employee devices and on-device AI interactions must therefore be discoverable. APIs, a primary means of AI interaction, must be continuously discovered as the new pace of coding drives 30% year over year growth in API attacks. Within your network, baseline traffic between applications, actors, and workloads. Beyond your four walls, continuously monitor for domain and brand abuse across social media and marketplaces, and scan the dark web for potential breaches.

Review Inventories: In the frontier LLM era, we continue to hear about “too many risks.” Contextualization and prioritization are imperative to avoid analysis paralysis. Inventories should cover:

• SBOMs and dependencies
• APIs, agents, applications, workloads, endpoints, and MCP servers
• certificates, domains, hostnames, and public-facing surfaces
• employees, customers, and workloads
• SDLC and DevOps: code gating and automated testing.
• Runtime: Web Application Firewalls, firewalls for new AI vectors (Akamai), and API, LLM, and MCP gateways that enforce authentication, authorization, and consumption limits to mitigate unbounded consumption (OWASP LLM06).
• Employee devices: agents that enforce approved AI tools and prevent sensitive data from being pasted into convenient AI clients.
• External surfaces: takedown processes with registrars, marketplaces, and social platforms.

Automated review against known risks, compliance and security frameworks (HIPAA, PCI, GDPR, MITRE, OWASP), and existing compensating controls lets organizations rationalize these inventories, confirm governance, and plan technical controls.

Evaluate Controls: Clear, defined governance is necessary for human and non-human actors across surfaces. Best practice is defense in depth, but policies must be defined and enforceable. Start with software-based microsegmentation, ideally with a positive security model. It mitigates lateral movement, drastically reduces risk from shadow AI and the overprovisioned agents well-intending employees create, and keeps policy maintenance feasible.

On that foundation, layer controls at every stage:

Adjust Controls: Eyes on glass cannot keep up with the pace of change. Within the SDLC, controls belong in automation pipelines. Elsewhere, controls should be automated within the tools themselves (self tuning firewalls, circuit breakers, automated policy engines) or managed by the new actor we’ve been defending ourselves from: agents. Governed effectively, agents accessing controls via MCP servers will progress us toward self-healing ecosystems.

Monitor & Feedback Loop: Monitoring production environments, the SDLC and CICD, and external surface scanners closes the loop. A self-healing ecosystem requires continuous context, and leading observability platforms already expose that data to security agents via MCP.

Combined, these elements let organizations build towards a self-discovering, self-governing, self-monitoring ecosystem. This will not happen overnight, but there are three ways to start building towards the DREAM.

First, level up your automation program today. Are you already using terraform and infrastructure as code? Have you included WAF and CDN onboarding at the end of your deployment? What else can you automate?

Second, start using these AI tools in your security program as soon as possible. It is difficult to conceptualize policies and governance without familiarity. Challenge your team to create agents in safe environments (containers, airgapped labs), share notes, and have fun with it. Vibe coding can be satisfying, yet clearly shows the threat of shadow AI.

Third, implement microsegmentation and Application Security Posture Management (ASPM). Microsegmentation is the single most important security control we have against agents (Akamai), covering the risks we cannot see nor think of. ASPM is the natural complement, contextualizing the risks we can see in our code and preventing risky or non compliant code from reaching production. When offensive LLMs find thousands of vulnerabilities in our surfaces, microsegmentation buys time, while ASPM lets us focus on the top priority and start digging out.

Finally, that health system may have been early with the request, but it is not alone. AI offers immense promise to organizations of all sizes, if we can mitigate the associated risks. A holistic framework lets the security team sleep well at night, so the organization can realize the DREAM that AI offers.