Security Architect. AI Builder.
Published Author.

Nassim Taleb argues that some systems get stronger from stress. Most security architectures are designed to resist stress. The interesting question is: can you design a security architecture that improves from controlled adversarial pressure?

Cyber Hormesis is a framework for building security systems that don't just survive attacks — they learn from them. Drawing from biological hormesis (where small doses of toxins strengthen organisms), the framework proposes that deliberate, controlled exposure to adversarial inputs makes security architectures more robust than purely defensive designs.

This has practical implications for AI security: instead of trying to prevent all adversarial inputs to AI systems (which is impossible), you design the system to metabolise adversarial inputs and strengthen its defences in response.

Key principles:

• Controlled stress exposure: red-teaming as a continuous practice, not an annual event
• Adaptive response mechanisms: security controls that reconfigure based on observed attack patterns
• Anti-fragile monitoring: detection systems that improve their sensitivity with each incident
• Graceful degradation by design: failure modes that preserve critical functions while sacrificing non-essential ones

The opposite of resilience isn't fragility. It's the illusion of resilience — systems that look robust until the first novel stress reveals they were merely rigid.

Taleb gave us the Black Swan: the unpredictable, high-impact event that no one saw coming. But in my experience, organisations rarely fail because of events no one could have predicted. They fail because of events everyone could see but no one acted on.

I call these Grey Ducks. They're not black swans (unpredictable) or white swans (expected and planned for). They're the visible, acknowledged, discussed-in-meetings risks that somehow never get resourced, never get mitigated, and never get escalated — until they cause exactly the damage everyone quietly knew they would.

Examples in cybersecurity:

• The legacy system everyone knows is unpatched but "too critical to touch"
• The third-party vendor whose security posture is "concerning" but whose contract renewal is "already signed"
• The AI deployment that "probably needs a governance framework" but "we'll get to it after launch"

The most dangerous phrase in any organisation: "We're aware of it."

AI systems drift. Not catastrophically, not overnight, but gradually — in ways that are invisible until the output no longer matches the intent. The Continuous Coherence Protocol is a monitoring and correction framework designed to detect and address coherence drift in AI systems operating in regulated environments.

The core insight: alignment isn't a one-time calibration. It's a continuous process that requires architectural support. CCP defines monitoring checkpoints, coherence metrics, and correction mechanisms that keep AI systems operating within their intended parameters over time.

The evolution to NS-CCP (Neuro-Symbolic Coherence Protocol) introduces symbolic reasoning constraints alongside statistical monitoring, enabling hybrid verification approaches for AI systems where pure statistical monitoring is insufficient.

Relevant in any environment where AI decisions have consequences that compound: financial services, healthcare, critical infrastructure, legal compliance.

Most AI implementations fail not because the model is wrong but because the model doesn't understand the organisation it's operating in. OCP is a structured approach for embedding organisational context — business rules, cultural norms, regulatory constraints, communication patterns — into AI system design.

Designed to bridge the gap between generic AI capabilities and organisation-specific deployment requirements.

Security architectures focused solely on prevention are incomplete. Cyber Resilience is a framework for designing systems that maintain critical operations during and after security incidents, with structured recovery paths and adaptive learning from each event.