You built a gate. The gate checks every tool call your AI agent makes. If the call violates policy, the gate denies it. The agent never reaches the tool. Clean architecture.

Now: what happens when the gate throws an exception?

If the answer is “the tool runs anyway,” you don’t have a gate. You have a suggestion box. And the agent is free to ignore suggestions.

This is the fail-open problem, and it is the single most common structural defect I see in AI agent security tooling. Not the most exotic. Not the most sophisticated. The most common. And it survives in production because it doesn’t show up in a demo, doesn’t trigger in a unit test, and doesn’t raise an alarm in a dashboard. The security layer fails silently, the agent continues happily, and nobody knows the gate was advisory all along.

Six weeks ago I published a post about moving from statistical detection to cryptographic proof for AI agent accountability. The project was called raucle-detect, it was licensed under AGPL-3.0, and it shipped the core capability receipt primitive: a signed, content-addressed record of every policy-gated action an AI agent takes.

A lot has changed since then. The project is now called raucle (not raucle-detect), it’s licensed under Apache-2.0 (not AGPL), and it has shipped four new modules that make those receipts work across organisational boundaries. The paper is under submission at IEEE Security & Privacy 2027. Microsoft accepted the upstream contribution into their Agent Governance Toolkit. And the whole thing went through nine rounds of adversarial security review that found three fail-open bugs I’m glad we caught before someone else did.

This post is the update.

A regulator has questions about a decision your agent made last quarter. A customer is in litigation because an AI-generated action cost them money. An internal auditor needs to certify that your agent did not act outside its authorised scope between two dates.

In each case, the same question: what cryptographic record proves that?

If your AI stack looks like most stacks in 2026, the honest answer is — nothing does. You have logs. Logs are claims, not evidence. You have a vendor saying their classifier flagged X% of bad prompts. A vendor assertion is not an audit artefact. You have screenshots and incident write-ups. Those don’t survive cross-examination.

This is the gap I want to talk about. And it’s why, after three months of work, I am open-sourcing the project I’ve been building to close it.

In January 2024, a team of researchers at Anthropic published a paper that should have changed how every organisation thinks about AI model security. The paper — Sleeper Agents: Training Deceptive Large Language Models That Persist Through Safety Training — demonstrated something uncomfortable: you can train a large language model to behave perfectly during evaluation and safety testing, while hiding a backdoor that activates only under specific conditions. And once that backdoor is in place, standard safety techniques don’t remove it.

Two years on, the implications are becoming harder to ignore. Organisations are integrating LLMs into critical workflows — code generation, decision support, customer interactions, financial analysis — often using open-weight models downloaded from public repositories. The sleeper agent research asks a question that most of those organisations haven’t answered: how would you know if the model you deployed was compromised before you ever touched it?

On 31 March 2026, Anthropic shipped version 2.1.88 of their Claude Code CLI to npm. Inside the package sat a 59.8 MB JavaScript source map file — a debugging artefact that mapped the minified production bundle back to the original TypeScript. That source map pointed to a publicly accessible zip archive on Anthropic’s Cloudflare R2 storage bucket. Within hours, the archive — roughly 1,900 TypeScript files and over 512,000 lines of code — had been downloaded, backed up, and forked more than 41,500 times on GitHub.

This was not a breach. No customer data was exposed. No credentials leaked. Anthropic called it “a release packaging issue caused by human error,” and that description is accurate. But the incident is worth examining closely, because it exposes a class of risk that most organisations building and shipping software at pace have not adequately addressed: what happens when your build pipeline becomes the vulnerability?

The chatbot answered your question. Annoying, maybe. Dangerous? Not really.

Now imagine this: you ask your AI assistant to book a flight. It reads your email, finds your passport number, accesses your corporate card, makes the reservation, and emails you the confirmation. All in thirty seconds. All without asking.

That’s an AI agent. And it represents the biggest shift in computer security since the cloud.

The Threat Model Flip

Traditional AI security focused on outputs. Could the AI generate harmful content? Could it reveal training data? Could it be jailbroken into saying something embarrassing? These were problems of persuasion—tricking an AI into convincing a human to do something bad.

The Incident Nobody Trained For

You’ve got a SOC. You’ve got runbooks. You’ve done tabletop exercises for ransomware, for data breaches, for insider threats. You’ve got your incident response team on speed dial.

But what happens when your AI customer support chatbot tells a customer they’re entitled to a refund they never made, cites a non-existent policy, and accidentally leaks another customer’s PII in the same conversation?

Is that in your playbook?

Building Trust Before Deployment

AI governance has a credibility problem. It sounds like compliance theatre—a box-ticking exercise—right up until your first model hallucination reaches a customer, an autonomous agent makes an irreversible business decision, or regulators ask questions you can’t answer.

The challenge isn’t that AI is ungovernable. The challenge I see is most organisations are trying to govern AI as an afterthought—a policy layer applied after systems are already in production. It’s like PCI compliance: checking what needs to be done when you’re already taking payments. That’s fundamentally backwards.

Post‑quantum cryptography has an image problem. It sounds like science fiction right up until it becomes an architecture constraint, a vendor question, and a programme of work you can’t avoid.

NIST finalised its first post‑quantum standards in August 2024—FIPS 203 (ML‑KEM), FIPS 204 (ML‑DSA), and FIPS 205 (SLH‑DSA). The algorithms are no longer theoretical. The migration work, however, has barely started for most organisations.

The risk isn’t a sudden quantum apocalypse. The risk is more mundane: cryptography is embedded everywhere, and changing it takes years.

Cloud security still gets framed as a configuration problem: open storage, overly permissive security groups, the usual “how did this bucket become public?” postmortem.

That’s not wrong. It’s just not the interesting part anymore.

The incidents that really hurt tend to be runtime stories: stolen credentials, abused roles, compromised workloads, and quiet data access through legitimate APIs. In cloud environments, the control plane is the battlefield. And the control plane speaks identity.

Supply chain security has a habit of refusing to fade, and it’s not because the industry enjoys pain. It’s because the attacker’s logic is perfect: why break into production when you can get your pipeline to deliver the compromise for you?

That’s not a theoretical risk. It’s a simple inversion of trust.

Every security trend eventually circles back to identity, because identity is how modern systems decide what’s allowed to happen. That’s true in cloud. It’s true in SaaS. It’s true in DevOps. And it’s becoming brutally true in AI.

The twist is that identity isn’t mostly people anymore.

The first wave of generative AI security was obsessed with content. Hallucinations. Toxicity. Brand risk. What the model might say if you asked it the wrong thing, the wrong way.

The second wave is about something more dangerous: capability.

Once an LLM can call tools—open tickets, query internal systems, send emails, push code, trigger workflows—it stops being a chat interface and starts looking like a new kind of privileged workload. Not malicious by design. Not even unreliable in the way people assume. Just connected to real systems with real permissions. And that is enough.

As we’ve moved through the second half of 2025, I’ve found myself spending a lot more time in the weeds of Industrial Control System threats. Not because it’s trendy, but because the shape of the modern supply chain is changing so quickly it’s hard to ignore. Next-gen automation, AI-driven decision-making, digital twins, and edge compute are turning supply chains into something closer to a nervous system than a chain—signals flowing constantly between software and the physical world.

If AI security still feels like a chaotic mix of research papers, vendor promises, and scattered best practice, that’s because—until recently—it lacked a shared vocabulary. In “normal” cybersecurity, that vocabulary has existed for years. When someone says “initial access” or “credential dumping”, a whole chain of assumptions and countermeasures snaps into place. In AI and ML, those conversations have often been fuzzier, full of hand-wavy phrases like “model manipulation” or “data poisoning” without a consistent way to describe what’s actually happening.

That’s why MITRE ATLAS matters. It’s not magic. It doesn’t secure anything on its own. But it gives defenders something the AI space has badly needed: a structured way to talk about adversarial behaviour against AI systems, grounded in real techniques and mapped to mitigations that can be engineered, tested, and improved over time.

The cybersecurity landscape has shifted in a way that’s easy to miss if you’re still looking for the usual signatures: misconfigurations, missing patches, careless access control. AI hasn’t replaced those problems. It has simply added a new layer where the failure modes aren’t always obvious, and where an attacker doesn’t need to “break in” if they can influence what the model learns or how it decides.

AI is no longer a skunkworks experiment running in a separate corner of the organisation. It’s being embedded directly into fraud detection, forecasting, customer support, quality assurance, routing and logistics, security analytics, and decision-making workflows that carry real commercial weight. That changes what “risk” looks like. It also changes what threat modelling needs to cover. The demand for threat modelling and risk assessments tailored to AI workloads is rising fast for a simple reason: traditional approaches, while foundational, often under-describe what can go wrong when the system’s behaviour is learned rather than explicitly coded.

Across multiple industries there’s a very real shift happening: AI isn’t just another line item on the risk register anymore, it’s the thing keeping security leaders awake at night. Talk to CISOs right now and the same theme surfaces again and again — the World Economic Forum’s Global Cybersecurity Outlook 2025 captured it well: Artificial Intelligence and Large Language Models (LLMs) have overtaken ransomware as the top concern in many boardrooms.

Federated Learning (FL) is one of those ideas that sounds almost too convenient when you first hear it. “Train a model across lots of organisations, but don’t move the data.” In a world where data is radioactive—healthcare records, financial histories, anything covered by regulation or common sense—that’s an enticing promise.

And it’s a real shift. Instead of dragging sensitive datasets into a central lake and hoping governance keeps up, you ship the learning out to where the data already lives. Each participant trains locally, then shares model updates—typically gradients or gradient-derived weight deltas, not the full learned parameters—back to a coordinator. The raw data stays put. On paper, everyone wins: better models, better privacy, fewer legal headaches. The foundational architecture was laid out by McMahan et al. in their 2017 paper on communication-efficient learning from decentralised data, and the field has moved quickly since.

Federated Learning Security: Training Together, Staying Safe

Federated Learning is one of those ideas that sounds almost too convenient when you first hear it. “Train a model across lots of organisations, but don’t move the data.” In a world where data is radioactive—healthcare records, financial histories, anything covered by regulation or common sense—that’s an enticing promise.

And it’s a real shift. Instead of dragging sensitive datasets into a central lake and hoping governance keeps up, you distribute the model out to where the data already lives. Each participant trains locally, then shares model updates—the learned parameters—back to a coordinator. The raw data stays put. On paper, everyone wins: better models, better privacy, fewer legal headaches.

But there’s a catch. There’s always a catch.

Edge AI is one of the most exciting shifts in modern architecture, not because it’s new, but because it’s finally usable. The pitch is simple: move intelligence closer to where data is created, reduce latency, keep sensitive information local, and stop treating connectivity like a guarantee. For industrial systems, retail analytics, logistics, smart environments, and a wide range of sensor-heavy use cases, that shift can be the difference between “interesting demo” and “operationally valuable”.

But it comes with a security cost, and it’s a familiar one. The moment you distribute capability, you distribute risk.

Edge AI isn’t just cloud AI deployed somewhere else. It’s cloud AI stripped down, quantised, compressed, and pushed onto devices that were never designed to behave like hardened servers. It’s also frequently deployed into environments where physical access is plausible. That changes the threat model entirely. The attacker isn’t always remote, and they don’t always need a zero-day. Sometimes they just need a screwdriver and a quiet moment.

Zero Trust for AI: Securing Intelligence in a Distributed World

Traditional perimeter security has been dying for years. AI just accelerates the funeral.

The reason is simple: modern AI isn’t a single application sitting behind a firewall. It’s a distributed system of systems—data pipelines, feature stores, training jobs, model registries, inference endpoints, retrieval services, agent tools, serverless glue, observability layers—often spread across cloud accounts, regions, vendors, and increasingly, edge devices. In that world the idea of a “trusted internal network” stops being a control and starts being a comforting story.

If an attacker gets a foothold anywhere inside that story—an over-permissioned service account, a compromised CI runner, a leaked token in a notebook, a poisoned dependency from a public model hub—the blast radius can be spectacular. Not because AI is magic, but because AI systems are connected, privileged, and hungry for data. Lateral movement becomes an architectural feature.

This is why Zero Trust isn’t a branding exercise for AI. It’s the minimum viable security stance for distributed intelligence. The core idea—never trust, always verify—dates back to Forrester’s original work and is formalised in NIST SP 800-207, but applying it to AI systems demands attention to components that traditional IT rarely considers: model weights, training data, inference pipelines, and agent toolchains.

Hardware-Enforced AI Security: Fortifying Your Models from the Ground Up

I’ve spent a lot of time thinking about the layers of defence we keep piling onto AI systems—secure coding, locked-down pipelines, adversarial testing, red-teaming, the whole lot. All of it matters. But there’s a blunt truth underneath the lot of it: if the underlying hardware and execution environment can’t be trusted, you’re ultimately building a very clever house on sand.

Adversarial Robustness Pipelines: Building AI That Bends, But Doesn’t Break

AI has properly arrived. I use models on my laptop daily, I’ve got agents digging up research for me across the internet, and half the tools I use now have some flavour of “intelligence” baked in. But, as tends to happen, whenever we invent a new way to do something clever, someone else invents a new way to break it.

Securing MACH Architectures

The business landscape is fiercely competitive, and digital innovation has become a baseline expectation. Businesses need agility, scalability, and solid security to keep pace. One architectural approach that has taken hold is MACH (Microservices, API-First, Cloud-Native, Headless). I’d seen hints of it in large organisations wrestling with legacy systems, but it was only when I helped a client build an entire banking infrastructure from scratch in the cloud that I really understood what MACH makes possible – rapid adaptation when competition and technology are both moving fast.

In Search of a Secure Mobile Phone

The older you get in security, the less you believe in absolutes. “Secure” becomes a moving target. “Private” becomes a trade-off. And the smartphone — this glowing slab that knows where you are, who you talk to, what you read, what you buy, and what you think — is where those trade-offs become uncomfortably personal.

As an iPhone user, I’ve had my share of friction with Apple’s approach to data protection. When Advanced Data Protection arrived, I set out to enable it properly and ended up on a year-long detour through Apple Support that culminated in creating a brand-new iCloud account and effectively rebuilding my digital life from scratch. The experience wasn’t just annoying; it was revealing. Even when the platform offers stronger protections, the path to using them can feel like a negotiation with your own convenience.

Designing zero trust architectures for clients running thousands of IoT devices changes how you think about “the network”. These aren’t laptops and servers sitting behind a predictable boundary. They’re sensors tracking warehouse capacity, systems automating movement of stock, retail analytics devices counting footfall and interaction — tiny computers scattered across physical spaces, often deployed faster than they can be governed.

The problem isn’t that IoT is inherently insecure. It’s that it’s inherently uneven. You’re dealing with a mixed ecosystem of manufacturers, firmware baselines, protocols, and operational ownership, and then you’re expected to treat the whole thing like a coherent fleet.

Microsegmentation has become a key technique for improving security by isolating workloads and reducing the attack surface. Unlike traditional network segmentation methods, which carve up environments into broad zones based on subnets or VLANs, microsegmentation pushes the boundary down to where modern breaches actually play out: between individual workloads, services, and identities — the east-west traffic that perimeter defences never see.

With the industry’s attention locked onto AI, it’s easy to forget that another computing shift has been accelerating in the background. Quantum computing used to live comfortably in the “interesting theory, far from production” bucket. It doesn’t anymore. The serious money, the serious lab time, and the steady cadence of capability gains have pushed it into a category security architects can’t ignore, because cryptography is a foundational dependency, and foundational dependencies fail loudly when the underlying assumptions change.

Beyond Bitcoin: Blockchain’s Role in Next-Gen Security Architectures

Following on from our recent deep dives into cloud and container security, this post looks at a technology that, while often synonymous with digital currencies, has broader implications for security: blockchain. Its potential extends well beyond cryptocurrencies, with its decentralised and immutable nature making it a practical option for securing various applications, particularly in supply chain management and identity verification.

Securing the Supply Chain

Supply chain management is a complex process involving many stakeholders, each contributing to the journey of goods from production to consumption. Traditionally, this process has relied heavily on centralised databases — a clear vulnerability, making them targets for tampering and fraud. Blockchain offers a compelling alternative. By providing a decentralised ledger where every transaction is recorded immutably, it introduces a level of auditability that centralised systems struggle to match.

Serverless architectures sell a very particular dream: your code runs when it needs to, scales when demand arrives, and costs you less because you stop paying for idle capacity. It’s a persuasive pitch, and in many cases it’s true. But the most profound change serverless brings isn’t cost or scaling. It’s a security boundary shift.

In a traditional world, security teams built mental models around hosts. Patch the operating system. Harden the image. Control inbound ports. Monitor the box. In serverless — particularly Function-as-a-Service — that mental model collapses, because the “box” is not something you own, and in many cases it barely exists long enough to be observed in the way we’re used to.

Serverless does genuinely eliminate certain attack classes. There is no SSH to brute-force, no persistent operating system to rootkit, no long-lived process to hook into. That’s a real win. But the risks don’t disappear — they move, and the new risks are less familiar to most teams.

AI and LLMs are transforming how organisations build products and run operations. They’re also changing what “an incident” looks like. When a traditional system is compromised, you often get familiar signals: malware, lateral movement, suspicious binaries, noisy command-and-control (C2) traffic. With AI systems, the first sign can be quieter, more ambiguous, and far more dangerous to ignore.

Not a service outage. Not an obvious intrusion. Just a model that starts behaving differently.

Many security professionals feel comfortable with classic incident response but get uneasy when the discussion shifts to production AI. That unease is rational. Attackers don’t need to take your platform down to harm you. They can steer it. They can degrade it. They can manipulate it in ways that look like “business variance” rather than a cyber event — until the impact is real. The attack taxonomy is broader than most teams expect: data poisoning, model evasion, model extraction, and prompt injection are all distinct threat categories that require different detection and response approaches.

This post is about responding to incidents in AI systems: what to look for, how to reason about the signals, and how to organise a response when the system is technically healthy but operationally compromised.

In the last post I focused on securing the model development pipeline — the supply chain that turns raw data into deployed behaviour. This time the topic is a quieter problem. It doesn’t show up neatly in most “Top 10” lists, but it keeps surfacing in real AI deployments, especially when models get embedded in high-consequence workflows.

It’s the inexplicability threat: the security risk created when you cannot reliably explain why a model did what it did.

Not “AI is complex” in the abstract. Not “deep learning is hard”. The practical version: when you can’t account for a decision, you can’t confidently detect manipulation, prove integrity, or recover quickly after something goes wrong.

If model inversion is the attack where you interrogate a system until it starts leaking what it learned, then pipeline compromise is the attack where you don’t bother interrogating the model at all. You simply change what gets built.

That’s the quiet truth about AI security: the model is only as trustworthy as the machinery that produces it. And the machinery is bigger than most people assume.

The model development pipeline is the full chain that turns raw data into deployed behaviour — collection, preprocessing, feature engineering, training, evaluation, packaging, deployment, and retraining. It looks like “engineering” on a diagram, but it behaves like a supply chain. It crosses teams. It crosses platforms. It crosses trust boundaries. And it’s full of artefacts attackers would love to control: datasets, labels, notebooks, feature stores, training code, weights, evaluation reports, container images, API keys, and deployment manifests.

Last time I discussed training data poisoning: the upstream attack where adversaries influence what a model learns by manipulating the dataset you train on. This time the threat flips direction. Instead of corrupting the input to training, the attacker interrogates the trained model itself.

Model inversion attacks aim to infer sensitive information about the data a model was trained on. In the worst cases that can mean reconstructing attributes of real people—health indicators, financial details, identifiers—or revealing statistically sensitive features about a dataset that was assumed to be private.

A quick clarification: “model inversion” in the strict academic sense (Fredrikson et al., 2015) refers to inferring sensitive attributes from model outputs. A related but distinct class of attack is training data extraction, where the goal is to recover verbatim training examples—particularly relevant for large language models (LLMs). In practice the two concerns overlap, and this post addresses both under the broader umbrella of privacy attacks against trained models.

The uncomfortable premise is simple: if a model internalises patterns from sensitive data, then a determined attacker may be able to tease those patterns back out through carefully chosen queries.

In the last post I explored prompt injection: the trick of turning “content” into “instructions” once an LLM is embedded inside a wider application. Training data poisoning is different. There’s no jailbreak moment, no single malicious prompt. Instead, the attack happens upstream, quietly, while you’re building the thing.

It’s the most uncomfortable kind of security problem because it targets the part of the system that’s hardest to reason about: the data you trust.

Large Language Models (LLMs) are being stitched into more and more products, quietly changing what “an interface” even means. After years in cybersecurity, it’s tempting to shrug and say: it’s software, it’s data — what’s new?

The uncomfortable answer is that the boundary between software and data is blurrier than we’re used to. In classic systems, untrusted input sits on one side of a parser and code sits on the other. With LLMs, natural language is both the user interface and — effectively — the control plane. That’s where prompt injection lives.

Generative AI security stops being a theoretical debate the moment you run an uncensored model and realise how quickly it will comply with the wrong kind of curiosity. Over the past few months, a lot of my own time has gone into building machine learning models for cybersecurity in a personal project, and the practical takeaway has been blunt: capability is easy to unlock; control is the hard part.

That’s the real engineering challenge with modern language models. The risk isn’t simply that they can write code. It’s that they can write code at speed, at scale, and with a confidence that looks persuasive even when it’s wrong. For security engineers, this is not just “another developer tool”. It’s a new production dependency with a new class of failure modes.

IAM Roles for ECS/EKS: Right Permissions, Right Place

IAM role configuration is central to container security in AWS. The principle of least privilege applies everywhere, but it gets ignored in container orchestration more often than almost anywhere else. A common pattern: the infrastructure is locked down tight, but the container task runs with AdministratorAccess because someone couldn’t get S3 writes working late on a Friday. That’s a serious security gap worth closing.

Securing Container Images: Best Practices for a Robust Containerised Environment

Throughout 2024, my blog posts will draw on my security engineering and architecture experience, sharing practices and lessons from working with AWS over the past decade.

This month, I’m covering container security — specifically, how to protect your containerised infrastructure.

The three pillars I lean on for container images — and have applied in both financial services and government — are: Image Scanning, Immutable Infrastructure, and Signing with Verification.

Throughout my journey in different organisations over the past two decades, one thing has stayed stubbornly consistent: the decisive factor in cybersecurity is rarely the tooling. It’s culture. You can buy the best controls on the market, build immaculate architectures, and hang your walls with policies and certifications… and still get flattened because the organisation behaves in a way that makes security impossible to sustain.

That’s not a dig at people, either. It’s just reality. Security is a human system running inside a business system, and the human system has moods, habits, pressure, fatigue, and deadlines. If you want resilience — the kind where the organisation bends when it gets hit and doesn’t snap — then you’re not just designing controls. You’re shaping behaviour.

Penetration testing is a critical part of any robust cybersecurity strategy. But a “successful” penetration test shouldn’t be judged only by what the testers found; it should also be judged by what the organisation detected, how quickly it made sense of what it saw, and whether the monitoring stack told a coherent story of what happened.

That’s where the SOC comes in.

Being in a leadership role in information security requires you to hold an odd mix of things in your head at the same time: enough technical depth to smell nonsense, enough strategic thinking to steer the ship, enough leadership to keep people moving in the same direction, and enough business and communication skill to make any of it land with the people who control budgets and priorities.

That’s a lot. And it’s on top of the day job: projects, incident noise, risk decisions, the “hot topic of the day”, and the never-ending parade of meetings that all feel urgent right up until they collide with each other in your calendar.

Over many client engagements I’ve tried all sorts of methods to keep myself from spinning out — some fancy, some simple, some frankly a waste of time. What follows is what I still use today because it actually works in the real world, not just in an executive coaching book.

The ability to communicate effectively with a diverse array of stakeholders — from your own security team to C-level executives — is likely the single most important aspect of security leadership. It is where technical expertise meets business reality, and getting it wrong usually means failing to secure the budget, support, or cultural buy-in you need.

Striking the right balance is an art form. Over the years, I have been fortunate enough to hold leadership roles across different organisations, and I have had mentors who drilled one lesson into me above all else: when you are talking to the C-suite, stop talking about technology and start talking about risk.

As I get ready to step into my next assignment, I keep coming back to the same truth: the first couple of weeks in a CISO role can make or break the next twelve months.

Not because you’ll “fix security” in a fortnight. You won’t. Nobody does. But because those first days are when you learn what you’re really walking into — how the organisation behaves under pressure, where the trust sits (and where it doesn’t), and whether the security team is seen as a partner or as the people who turn up late and say no.

Working as an independent consultant gives you a strange advantage: you get to watch the same organisational patterns repeat across completely different industries. New leadership arrives. A new strategy deck appears. A familiar set of “transformations” roll through. And somewhere in the noise, the security team is expected to keep the lights on, keep the auditors happy, and somehow also “be more proactive” against a threat landscape that changes faster than most businesses can plan.

This came up recently in a catch-up with a mentee. Their organisation was going through leadership change, and the security function was feeling the same gravitational pull it always does in those moments: everything becomes urgent, everything becomes visible, and everything becomes a priority.

Which brings us to a syndrome that turns up again and again, especially with new managers trying to prove they’re decisive: “Everything is Priority One”.

Third-Party Risk Assessment: Why Your Security’s Only as Strong as Your Weakest Vendor

I’ve been doing this security architecture gig long enough to spot the pattern. You spend a fortune building a digital fortress. You patch everything, you train your staff until they’re sick of your voice, and you deploy Zero Trust architectures that would make a bank jealous. You feel good. You feel secure.

Then, inevitably, the phone rings. It’s not your firewall that failed. It’s not your sophisticated intrusion detection system that missed a beat. It’s the bloke who prints your shipping labels, or the survey platform marketing signed up for with a credit card, or the facilities management company that has remote access to your HVAC system. They got popped, and now your data is floating around the darker corners of the internet.

AI in Cybersecurity: When the Shield Becomes a Sword

You can’t open a browser tab these days without someone banging on about how AI is going to either save humanity or turn us all into paperclips. It is everywhere. And in our corner of the universe—cybersecurity—it has created this properly fascinating, if slightly terrifying, dynamic that is fundamentally changing how we think about defence.

As an architect, I look at AI and Machine Learning and I see enormous potential. But I’d be lying if I said I didn’t also see a massive headache brewing. Because here is the uncomfortable truth we don’t talk about enough: the exact same technology that promises to help us predict threats before they materialise is simultaneously giving attackers a significantly bigger stick to beat us with. We are essentially engaged in an arms race where both sides are wielding the same weapons. The question isn’t whether AI will transform cybersecurity—it already has—but whether we can stay ahead of the curve.

As a security leader, one of the most important parts of the job has never been the technology on its own. The technical work matters—but what really decides whether security succeeds or becomes theatre is the team behind it. Not the org chart, not the tooling stack, not the maturity model slides. The actual humans who show up, carry the pressure, make the judgement calls, and keep going when everyone else only notices security during a crisis.

We say it all the time: your cyber defences are only as strong as the people safeguarding them. It’s true, but it’s also a bit of a cop-out, because the unspoken bit is this: if you want a strong security function, you have to build the conditions for it. You have to shape the team, protect it from nonsense, and give it room to grow. That’s not fluffy leadership talk—it’s operational reality.

The task of managing cybersecurity risks in supply chain management isn’t just “nice-to-have” anymore. It’s basic survival. Modern supply chains aren’t a neat line from supplier to factory to customer; they’re a living network of companies, contractors, platforms, software components, and outsourced services, all stitched together with APIs, portals, shared data, and a lot of trust.

And trust, in security, is a dangerous thing to hand out by default.

A typical day tells the story.

You wake up and reach for the phone beside your bed, thumb through notifications, skim the news, and half-consciously accept that your morning begins inside someone else’s platform. Maybe you order a flat white through an app. Maybe you call someone overseas. Maybe you open your email and step into work. None of it feels remarkable anymore, because the modern web has done its best work when it feels invisible.

Securing our businesses from persistent, adaptable adversaries is an imperative. It requires orchestration that looks, from a distance, like a symphony: people, process, and technology working together with the timing and discipline to detect quiet signals early, make sense of them quickly, and respond without chaos. The challenge is that most organisations try to build this capability while facing a scarcity of skilled personnel, inconsistent funding, and a culture that only pays attention when something has already gone wrong.

In the French Alps, the weekend ritual is as predictable as the weather isn’t. You see the same choreography at trailheads and lift queues: people tightening boots, checking bindings, shouldering packs, and doing the quiet mental arithmetic of risk. Ropes. Layers. Food. A head torch “just in case”. And increasingly, a small constellation of electronics that promise to make the mountain feel a little more knowable.

A watch that tracks altitude and heart rate. A phone full of offline maps. A satellite messenger clipped to a strap. A beacon, dormant until the worst day of someone’s life. These devices are sold as safety, and often they are. But they also represent something else—something we don’t tend to think about in the cold wind, with gloves on and a ridgeline ahead.

They’re computers. Networked computers. And the mountain doesn’t make them less hackable.

Digital business operations continue to expand, and the threat landscape evolves in lockstep—more complex, more professional, and more opportunistic. Attackers are no longer “finding vulnerabilities” in the abstract; they’re running an ecosystem. They share tooling, reuse techniques, buy access, and iterate faster than many internal teams can patch. In that context, the question isn’t whether threats exist. It’s whether an organisation is forced to learn about them only after impact.

That’s where threat intelligence earns its place.

Security is easiest to add when nothing has been built yet.

That sounds obvious, but it’s the most expensive lesson teams learn the hard way. Once a system is deployed, every “security fix” becomes a negotiation with time, budget, and inertia. In the design phase, it can be a decision. That’s what people mean by secure by design, and threat modelling is one of the few practices that makes that phrase real.