The Defensive Donut: Why AI Needs a Ring of Security to Stay Safe

The Defensive Donut: Why AI Needs a Ring of Security to Stay Safe

By ICTpost Tech Desk

Securing AI with a Donut of Defense: A Smart, Layered Approach to Trustworthy Intelligence

Artificial Intelligence (AI) is no longer an experimental frontier — it’s at the center of modern digital life. From financial forecasting to disease diagnosis, from customer service chatbots to autonomous vehicles, AI is powering mission-critical systems around the world.

But here’s the problem: while AI stands tall at the center, its perimeter often lacks proper defenses. It’s a high-value target with wide-open access points. As AI expands, so do its risks — from misuse and data leakage to outright manipulation and cyberattacks.

The Defensive Donut Model

Cybersecurity experts now propose a straightforward framework: wrap your AI in a “defensive donut” — a ring of protection composed of four layers: Discover, Assess, Control, and Report.

1. Discover: Shine a Light on All AI Activities

“You can’t secure what you can’t see.”

The first step is to detect every instance of AI usage across your organization — both authorized and unauthorized.

The Rise of Shadow AI

  • A 2023 Cisco Security Report found that 74% of employees admitted to using AI tools like ChatGPT at work without formal approval.
  • These unauthorized tools — called Shadow AI — can introduce risks like data leakage, poor model performance, or compliance violations.

What to Discover

  • AI models being used (e.g., OpenAI’s GPT, Google’s PaLM, Hugging Face models).
  • Platforms running AI (on-premises, AWS, Azure, Google Cloud).
  • APIs and integrations consuming AI-generated outputs.

Tools like Microsoft Purview and IBM Watson OpenScale help detect and manage AI assets across large environments.

2. Assess: Evaluate Vulnerabilities and AI Posture

Once discovered, each AI system must be evaluated for security gaps — a process known as AI Security Posture Management (AISPM).

Key Assessments:

  • Vulnerability scanning of AI models and surrounding infrastructure.
  • Penetration testing (or red teaming) to simulate real-world attacks.
  • Third-party model inspection for malware or backdoors.

In May 2024, SecurityWeek reported that attackers compromised a Fortune 500 company’s chatbot by uploading a tainted LLM from an open-source repository. The model contained prompt injection backdoors to exfiltrate user prompts.

Hugging Face hosts over 1.5 million models, many with little to no vetting. According to a 2024 MITRE Study on AI Threats, 47% of open-source models carried at least one known vulnerability.

Organizations should scan AI artifacts with tools like HiddenLayer and align with standards such as:

3. Control: Block, Monitor, and Guide AI Interactions

Once vulnerabilities are addressed, controls must be implemented to prevent new threats — both from users and external actors.

a) AI Gateway

An AI gateway monitors and filters prompt inputs and outputs.

Top Risk: Prompt Injection

  • Listed as the #1 threat in the OWASP Top 10 for LLMs.
  • A form of social engineering that tricks AI into violating instructions or security constraints.
  • Example: Researchers in 2023 demonstrated prompt injection techniques that bypassed OpenAI’s filters using simple wordplay.

Mitigation strategies:

  • Input/output sanitization
  • Prompt validation and contextual filters
  • Rate limiting and session management

b) Privacy and Data Controls

It’s not just about what goes into the AI — it’s also about what comes out.

In April 2023, Samsung employees inadvertently leaked confidential data to ChatGPT, resulting in an internal ban on AI use.

Controls Needed:

  • Data Loss Prevention (DLP) systems
  • Output monitoring tools
  • Internal fine-tuning to align models with ethical and regulatory requirements

4. Report: Visualize, Govern, and Comply

Security must be visible and reportable.

a) Dashboards for Risk Prioritization

Dashboards should provide:

  • Real-time threat analytics
  • AI usage heatmaps
  • Trends in misuse or misconfiguration

Example: Palo Alto Cortex XSIAM includes AI-aware dashboards and threat detection tools for monitoring LLM usage.

b) Regulatory Compliance

Laws that govern AI use and data include:

Compliance can be managed with tools mapped to:

AI at the Center, Security All Around

The “Defensive Donut” is more than a metaphor — it’s a blueprint for secure AI. When you implement the four layers — Discover, Assess, Control, Report — you turn AI into a safe, strategic advantage.

AI may be the future, but without security, it’s a future full of risk.

Key Takeaways:

  • 74% of employees use unauthorized AI tools — Cisco, 2023
  • Hugging Face hosts over 1.5 million modelssource
  • Prompt injection is the #1 threat to LLMs — OWASP, 2023
  • Regulatory guardrails coming fast — GDPR, DPDP, EU AI Act

For more stories on AI, cybersecurity, and emerging tech in India, stay with ICTpost.com – where innovation meets insight. editor@ictpost.com

Did you like this? Share it:

Leave a Reply

Your email address will not be published.

69  +    =  79