Building a Black Belt Purple Team: AI, Shadow IT & Cyber Resilience
🎧 Listen now to The Resilience Room Ask most people how they ended up in cyber security and you'll get a story about a plan. Ask Kev Marriott and you'll get a story about happenstance. Kev spent the best part of two decades at British Telecom. He started in sales, became a trainer and an NVQ assessor, then drifted toward IT incident management — keeping the network, the platform and bt.com live and reducing downtime wherever he could. There's a lot of crossover between network operations and security, and over the years that crossover pulled him in. He ended up running the incident, problem and change management team for BT, coordinating the response to large-scale cyber and physical security incidents across critical national infrastructure. Which also meant owning the playbooks. And being the person on the end of the phone at four in the morning. "You have the joy of being the person called out at four o'clock in the morning if something goes wrong, to make the decision of what do we do. Do we stand something up? Do we get the team together? Do we reach out to third parties?" What turned the job into a passion was seeing the other end of it. Before moving into security, Kev answered calls from customers who had lost savings to fraud — people whose lives had been materially damaged by an attack somewhere upstream. That's the thing that stuck. Key takeaways Purple teaming is a culture, not a headcount. Some mature organisations do appoint someone to lead it, but the value comes from collaboration between red and blue, not from an org chart change. The gap between compromise and exfiltration keeps shrinking. Triage windows that were comfortable a few years ago are already too slow. Shadow AI is shadow IT's successor — and it arrives with a much bigger data problem attached. Your supply chain's AI is now your problem too. Without data lineage, you can't answer whether your model has been poisoned. Governance and good hygiene haven't been replaced by AI. Identity and access management, critical asset lists and network topology matter more now, not less. What purple teaming actually is The mechanics are straightforward. A red team gets sign-off from the CISO to test a defined part of the estate, with a scope and a set of goals they don't stray outside. The blue team usually knows nothing about it — deliberately, so that when they do spot something, they respond as though it's a genuine attacker. Then, at some point, the red team is called in and the two sides sit down together. That sit-down is the purple team engagement. And Kev's point is that it isn't a formality — it's where the baseline actually moves. "Blue teams are incredible. But the attack team always has the first move advantage." What surprised me in the conversation was how little of this is technical. Purple teaming is mostly a social and organisational problem: getting two groups who are structurally set against each other into a room, regularly, with a shared interest in the outcome. Kev is honest about the friction. Red team engagements can run for weeks. If the blue team finally spots something on Christmas Eve and gets called out, that doesn't build warmth between departments. The answer isn't to remove the competition — it's to make it healthy. "Competition drives things forward. Every time a red team is thwarted because the blue team has learned something, you'd hope the red team takes a bit of a front to that and goes, well, you blocked that, but now I'm going after this." The real unlock is being able to ask why. When a red team makes an odd jump between network boxes, a blue team looking at the topology can be genuinely baffled by the choice. In a purple team environment, you can simply ask them. Sometimes the answer is "we were just sniffing around." But over time, those conversations shape which boxes you forensically capture first, and which you prioritise for review. That's a process improvement you can't buy. Kev returns to curiosity repeatedly — it's the first thing he looks for when hiring. Not what someone already knows, but whether they'll keep asking why. Shadow AI, and a timeline that keeps contracting The industry never quite solved shadow IT. It improved, but it didn't go away. Now it has a successor. Shadow AI is people bringing the tool they use at home into work because it's familiar — sitting outside your governance model, outside your guardrails, ingesting whatever gets pasted into it. The consequences run from GDPR exposure to unreleased financial results leaving the building. "The cyber industry was built on 'never knowingly give your data anywhere else.' AI is almost the antithesis of that. It's very much, give us all your data, as much as you can." That's the genuinely new thing, in Kev's view. We spent years building behaviours — don't click the link, don't trust the email from the CEO, don't hand over your data because someone asked. AI risks quietly reversing some of that progress. Meanwhile the clock is running faster. Kev deliberately wouldn't quote specific numbers ("even by the time this podcast is launched, it would probably be out of date"), but the shape is clear: a triage window that a SOC might reasonably have taken half an hour over a few years ago is now too slow. They're in and out before you've opened the ticket. Immersive's purple team work in this space is being built against MITRE ATLAS, showing both sides of the same attack — how you'd run it, and how you'd spot, mitigate and prevent it. The same logic shows up in vulnerability management. Kev pointed to Mythos, Anthropic's AI system for autonomous vulnerability discovery, as a conversation-starter — and was careful to frame this as his own view rather than Immersive's. Whatever you make of the tool itself, the point stands: vulnerability management has historically been organised around zero-days, while proficient threat actors have long chained together lower-severity flaws to get in and back out again. If discovery is about to happen at machine speed, you need SLAs and processes for every vulnerability across software, firmware and hardware — not just the headline ones. Everything old is load-bearing again None of this works without the unglamorous stuff. Identity and access management, now for agents as well as humans. Critical asset lists. Knowing your topology and how data flows through it. Knowing who owns an application and how quickly it can be patched. Then there's the supply chain. Assurance used to mean asking a vendor about their patching schedule and incident response process. Now you also need to know about their AI — because if you're exchanging data with them and you have no data lineage, you can't answer whether your model has been poisoned. And there's the money. Kev raised a question I hadn't considered: where does the AI budget actually sit, and what happens if you've reduced headcount on the assumption AI covers it, then burn through your tokens? With providers shifting from subscriptions toward usage-based pricing, ROI and financial risk end up as the same conversation — which is exactly the kind of conversation that forces security, IT, finance and the board into the same room. Which, as I put it slightly flippantly during the episode, is a rainbow team. Kev's verdict: unhelpful as a job title, but not wrong about where this is heading. The belts on the wall I couldn't let the episode end without asking about the row of martial arts belts behind Kev on camera. He trains in MMA three times a week — three and a half years in, and roughly nine months out from his black belt. He took it up later than most of the people he spars with. It turned out to be the most useful analogy in the conversation. Playbooks matter. You need them, and audits require them. But nobody woken at 4am is turning to page 47 to find their response. What carries you is experience, previous learnings, and muscle memory built through repetition — doing the same thing over and over, in slightly different ways, looking for the improvement each time. "You idealise the playbook as a result of the muscle memory you build and the evidence you capture while performing the activities — rather than relying on the playbook to guide you when you hit that point." It cuts both ways in a security team. Some people don't feel empowered to make a decision under pressure. Others feel a bit too empowered, and make calls they weren't authorised to make. Rehearsal is what calibrates that. Kev's closing note on AI is a fair one to end on. He's not an AI sceptic — he's clear about how much good it can do, particularly in automating the laborious work so blue teams can actually go threat hunting or talk to a part of the business they never had time for. "Understand the use cases. Get your basic principles in place, test it, assure it, secure it, and you're in a really good place. Until it all changes in a week." Over to you: Kev argues the biggest barrier to purple teaming isn't tooling — it's getting red and blue in a room often enough to build trust. If you've tried it, what actually made the collaboration stick? And if you haven't, what's stopping you? Listen to the full episode of The Resilience Room wherever you get your podcasts.It's Software, Not Magic: Navigating the Vulnerability Speed Wave and Shadow AI
🎧 Listen now to The Resilience Room What does a 31-year career at one company look like in cybersecurity? For Lee Stephens, it looks like research, marketing, sales, operations, and consultancy — and never a dull moment. In this episode of The Resilience Room, Lee sat down with host Sam Dickison to cut through the noise on ransomware, the collapsing patch window, Shadow AI, and the looming quantum reckoning. The Retail Ransomware Wake-Up Call Last year's wave of attacks on the UK retail sector made cybersecurity front-page news. For Lee, the most striking thing wasn't the sophistication of the attacks — it was the opposite. "What came home was actually the simplicity of the attacks. Those really fundamental, boring basics are absolutely critical — and involved in so many of the incidents we deal with." While the media spotlight has moved on to AI, the threat landscape hasn't changed nearly as much as the news cycle suggests. The fundamentals remain the same: strong unique passwords, kept systems, tested backups, and a plan for a bad day. As Lee puts it, you don't need to outrun the bear — just don't be the slowest runner. The Vulnerability Speed Wave: Your 90-Day Patch Window Is Gone Here's the stat that should reframe every conversation about patch management: 2021: Average time from CVE disclosure to active exploit — one year 2025: Shrunk to one month 2026: Now at one week and one day 2027 projection: Potentially an hour — or a minute This geometric progression is being driven by AI-assisted vulnerability research on both sides. The traditional 90-day patch window is functionally dead, and some vendors are already moving to fortnightly cycles. Speed is now a security control in its own right. AI in the SOC: Real Benefits, Realistic Limits SOC analyst burnout is real. The volume of alerts, the repetitive triage work — AI-assisted automation has genuine potential to help. But fully autonomous SOCs? Lee remains cautious. "At times AI feels like magic. It's not magic, it's software. And from thirty years ago when I did my computer science degree: garbage in, garbage out." His recommendation: don't chase the ten-times transformation. Chase the ten percent improvement and compound it. Organisations that have tried to automate everything often find they're spending more time maintaining automations than the work would have taken. Shadow AI: The New Shadow IT Problem Most people using unofficial AI tools aren't being malicious — they're trying to get their job done. But the risk is real, especially around feeding confidential data into models without proper data handling agreements. Some of Lee's clients require all data to remain in the UK at all times, which effectively rules out certain major tools that can't guarantee that under peak load. His cautionary scenario: a team uses AI to diagnose an infrastructure problem. The output gets translated for customers, then condensed for a senior exec. Each iteration drifts further from reality — and when the fix doesn't work, multiple conflicting versions of "the truth" are already in circulation. "AI has no conscience, it doesn't care. It's a prediction engine. It's software. Not magic — software." The solution is culture and training as much as technical controls, combined with IT teams designing better corporate solutions so people don't need to go off-piste in the first place. What's Next: Quantum's Y2K Moment Quantum computing has been "just five years away" for decades — but the trend lines are genuinely moving. More qubits, better coherence times, major advances every few months. The security implication is stark: the maths underpinning asymmetric encryption is trivially easy for a quantum computer to solve, which means every piece of encryption on the planet eventually needs replacing. NIST has published post-quantum cryptography standards, and vendors are beginning to implement them. But the migration is a massive undertaking. "It's akin to a Y2K moment where everybody needs to upgrade — for really no benefit at all, just to stay standard. The only disadvantage is we don't know when the moment is." The organisations best placed will be the ones who've already done their cryptographic inventory and know what they'd need to upgrade and in what order. Whether the topic is ransomware, patch windows, AI, or quantum — Lee's message is consistent: there's no silver bullet, no magic. Just fundamentals, applied consistently, with humans staying in the loop. After 31 years, it's the message that keeps working. The Resilience Room is hosted by Sam Dickison. New episodes explore the human and technical realities of cybersecurity with guests from across the industry.The Black Belt of Cybersecurity: What a Former Military Pilot Turned SecOps Leader Knows About Surviving the Chaos
🎧 Listen now to The Resilience Room In 2012, a young Air Force officer named Michael Vetri walked out of his commanding officer's office having just been told his career was about to change forever. He'd trained as a pilot. Now he was being reclassified into a field almost nobody in the room understood. "I'm so sorry you were reclassified into cyber," his flight instructor told him. "If you need someone to talk to, let me know." That's not a joke. That's how misunderstood cybersecurity was fourteen years ago — obscure enough that getting assigned to it sounded like a consolation prize, or worse, a diagnosis. Fast forward to today, and that reluctant recruit is Senior Director of SecOps at a SaaS company serving the biopharma industry, a black belt Krav Maga instructor, and — going by the enthusiasm with which he brought up Assassin's Creed unprompted — a certified nerd in the best sense of the word. Mike sat down on The Resilience Room and covered a lot of ground: emotional intelligence as a revenue driver, the four things that keep security teams from burning out, why AI is currently winning the fight, and how disarming a training knife in a Krav Maga studio taught him more about incident response than most corporate leadership seminars ever did. If you work in this field, several of these are going to land uncomfortably close to home. Good. That's the point. Empathy isn't a soft skill. It's a spreadsheet line item. Mike leads with rank, but not with authority — a distinction the military drilled into him early. "In the civilian world, you can't just say, 'I'm a captain, you're an NCO, you'll do what I say.' No — it's a lot more influential leadership, invoking those empathetic muscles more." He points to a Harvard Business Review study of 189 companies that measured how deeply organizations embedded emotional intelligence principles — self-awareness, self-regulation, empathy, motivation — into their culture. The companies that did it well beat their revenue targets by roughly 20% year over year. The companies that didn't, missed by about the same margin. For an industry that still occasionally treats "soft skills" as a nice-to-have bolted onto technical competence, that's a hard number worth sitting with — especially if your last performance review mentioned "communication" in a slightly ominous tone. The four pillars that keep security teams from quitting Burnout in cybersecurity isn't news to anyone reading this. But Mike's framework for fighting it is refreshingly concrete. He tells every team he leads that people will stay if you give them four things — not in excess, just in moderation: Time. If someone's been grinding through late nights, notice. Let them come in later. Build in recovery. Money. Pay fairly for what they contribute — full stop. Recognition. Not just for wins. How you handle underperformance, professionally and without ambush, matters just as much. Development. Give people room to grow into who they're becoming, not just who they were hired to be. One detail worth stealing: when an underperformance conversation is looming, Mike doesn't open with the problem. He opens with "how are you doing? Everything okay outside of work?" — asked more than once if needed. People under pressure, especially newer employees still trying to prove themselves, often won't volunteer that something's wrong. They'll white-knuckle through it instead. A little space to be asked, rather than told, changes what people are willing to say. During one incident that dragged on for months, Mike built a shift schedule with staggered four-hour windows so his team could claw back rest without falling behind. He also borrowed a page from Air Force operational risk models: a short questionnaire — how much sleep did you get, how many consecutive night shifts, how long since you've touched this system — that produces a risk score. Score too high, and you talk to a duty officer before you're cleared to keep going. It's the kind of instrumented self-check most SOCs could use and almost none have. Be the eye of the hurricane Ask any incident responder what makes a bad night worse, and it's rarely the incident itself. It's the Slack DMs. The "hey, what's going on?" pings stacking up while you're still trying to pull logs. Mike's answer is to act as an umbrella for his team — absorbing pressure from above so the people doing the actual work can think. That starts with managing expectations upward immediately: this is fresh, I'm still learning it, I'll update you in an hour. One sentence, enormous operational space. It tells leadership "I've got this" while telling the incident team "go dark and go deep." Underneath that is a military framework Mike leans on constantly: the OODA loop — Observe, Orient, Decide, Act. In a live incident, the goal is to move through that loop faster than whatever you're up against. The moment you get stuck in "orient," unable to commit to a decision, you've lost the initiative. Everything else — the calm voice, the reassurance, the "breathe, I've got your cover" — exists to keep the team moving through the loop instead of freezing in it. And underneath that is the least glamorous ingredient of all: reps. "You wouldn't want a surgeon who hasn't performed one in three or four months operating on you," Mike says. Tabletop exercises, drilled relentlessly, are what makes calm possible when the real thing hits. AI is winning the speed war — for now This is where the conversation gets genuinely unsettling, in the way only a good SOC-lead story can. Mike cited threat intel research his own team pulled showing that traditional, manually written phishing emails land at about a 12% click rate. AI-generated phishing is landing at 54%. That's not an incremental shift — it's a different category of threat. His framing borrows from military history: "Just like the English were doing great with their longbowmen until they were met with cannons — we're doing great against traditional malware, traditional phishing, until we're met with AI-based malware, AI-based phishing." He pointed to Phantom Raven as a case study in the new normal — a highly polymorphic strain that reshuffles its own hash and callback domains the moment defenders try to contain it, sometimes riding in through a malformed open-source package nobody scrutinized closely enough. Mike's answer isn't to out-human the machines — it's to fight AI with AI, using it to accelerate investigation and triage so humans can make faster calls. But he draws a firm line at execution: "I still think it's a bad idea to let AI execute actions on your security stack." Not because it can't make good calls, but because when it's wrong, it's wrong at machine speed — locking down a developer's test script as if it were a live threat and taking a production deploy down with it. Assessment and evaluation, yes. Pulling the trigger, not yet — not until AI has earned the trust that comes from track record, the same way a junior analyst earns it. He's also thinking about a subtler insider-risk problem taking shape as advanced AI systems move into more sensitive testing and evaluation programs. His concern isn't the technology itself — it's what happens when knowledge of where that technology lives becomes valuable enough to bribe someone for. New capability, same old human vulnerability. Train at black belt level, or don't bother training at all Mike's been doing Krav Maga since a college roommate walked in mid-Assassin's Creed session and told him the game's weapon disarms were based on a real martial art taught nearby. He put the controller down, went and checked it out, and got hooked on his very first class — which happened to be weapon disarming. He made it to black belt in 2014, taught for years, and — genuinely one of the better anecdotes on the episode — met his wife through a self-defense class he was running as a side gig at New Mexico State. Her kicks, he reports, had "so much pepper behind them." The parallel he draws to cybersecurity is the episode's sharpest idea: you're only prepared to handle the level you've actually trained for. A yellow belt can handle a yellow belt. Theoretically, maybe an orange. Train your team only against script kiddies and spray-and-pray attackers, and script kiddies and spray-and-pray attackers are all you'll ever be ready for. The nation-state actor doesn't care what belt is hanging in your dojo. It shows up in the small stuff too — like treating every incident, however minor, as if you might have to defend your handling of it in court one day. Heavily documented, no matter how low-level it looks. "Don't give the opposing attorneys any reason to pierce your veil," as Mike puts it. Complacency doesn't announce itself. It just quietly lowers the belt you're actually prepared to fight at. The throughline of this whole conversation is one Mike states almost in passing but clearly means as a core operating principle: technical skill gets you in the room, but judgment, empathy, and relentless preparation are what keep your team — and your organization — standing when things go sideways. AI is changing the speed of the fight. It isn't changing who has to decide what to do about it. Catch the full conversation with Mike Vetri on The Resilience Room – available on most podcast platforms.Hacking AI: The Jade Puffer Attack, Prompt Injection, and the Dead Internet Theory
🎧 Listen now to The Resilience Room Featuring Glenn Wilkinson, co-founder & CEO of Agger Labs · Hosted by Sam Dickison Some guests bring a war story. Glenn Wilkinson brings a stage show. Early in this episode he tells us about Snoopy — a tool he built during his time at SensePost that quietly profiled audience members’ phones during a live theatre production about privacy. As people reached for the venue Wi-Fi, Snoopy pieced together names, home addresses and travel histories from thin air. At the finale, the show’s star turned it all back on the room from the stage: welcome, John from Boston — and here’s a picture of your house. Glenn still lights up recalling the actor who fronted it crossing the room to introduce himself — “you must be the hacker” — and being disarmingly kind about the whole thing. It’s a fitting way in, because everything Glenn does runs on the same instinct: this says I can’t. Let’s see about that. Anarchist or engineer? Glenn traces the instinct to a childhood in Zimbabwe, where a first computer was a precious, irreplaceable thing — break it and there was no replacement. He got bored of the hardware fast and fell for software instead: the demo games that shipped on magazine cover-CDs, locked to one level. With no way to buy the full versions, he taught himself to bypass the copy protection just to keep playing. Early reverse engineering, dressed up as a kid wanting more game. That, he says, is the whole thing in miniature. The hacker mindset has two parts: a love of tinkering and puzzles, and a healthy disdain for authority. Or, as he puts it: are you an anarchist or an engineer? It’s a distinction the show keeps landing on with different guests — IT is about understanding how something works; hacking is about understanding how something breaks. “This says access denied. That’s interesting. I wonder if I can get past that—or make it do something different.” Glenn Wilkinson These days that instinct has a business behind it. Glenn is co-founder and CEO of Agger Labs, an anti-ransomware startup born out of a simple frustration: when his team pulled apart the existing EDR and antivirus tools, none of them were much good at stopping novel ransomware. Their answer is a lightweight endpoint agent that runs entirely locally — no signatures, no cloud lookups — and judges every program by its behaviour. Because ransomware has to act like ransomware, the agent lays deception traps for it to trip over, which keeps false positives near zero. It was also engineered from day one to be near-impossible to tamper with; customers now ask it to stand guard over their EDR, backup and logging servers too. Like a little bodyguard, Glenn says. The first AI-orchestrated ransomware attack Which brings us to the reason this episode is so timely. Sysdig’s threat research team recently documented what’s being described as the first observed end-to-end agentic ransomware attack — a campaign given the excellent name Jade Puffer. Glenn’s honest first reaction? “Wow, that’s so cool” — before catching himself, the way a virologist might admire a nasty new strain before remembering whose side they’re on. The important nuance: the ransomware itself wasn’t magically “AI-powered.” What was new is that the intrusion — the reconnaissance, the break-in, the lateral movement, the escalation — appears to have been orchestrated by an AI model. “The important word,” Glenn stresses, “is orchestrated.” At the low end, that might be nothing more than someone proxying commands through a model with its guardrails stripped off, asking what should I do next? and pasting the answers into a live target. How did anyone know it was AI? It got sloppy. The logging was verbose, and the Python payloads were littered with comments that exposed the model’s own reasoning — the kind of housekeeping a seasoned human attacker would have scrubbed before deploying. Glenn suspects it may even have been botched: the encryption key possibly never made it back to the attacker, which raises the grim possibility it was a test, a ransomware-as-a-service dry run, or thinly veiled wiperware dressed up as ransomware. The part worth sitting with This cutting-edge, AI-driven attack got its foot in the door through a Langflow server carrying a roughly year-old, unpatched remote-code-execution vulnerability. Every new layer of technology just finds a new way to exploit the fundamentals. The black piano fallacy This is Glenn’s drum, and he beats it everywhere — including on BBC Breakfast, where his advice to eight million viewers was simply: get your basic cyber hygiene right. He borrows a line from his friend Dominic White to explain why people get this wrong. Picture walking down a dodgy street at night, worrying that a piano might fall on your head. Your risk register is skewed: you should be thinking about getting mugged, not about the one-in-a-million piano. Security teams do the same thing — they get excited and anxious about the exotic new threat while the attackers quietly walk through the front door the basics left open. So what does AI actually change? Not the playbook — the clock. The timelines defenders used to enjoy (days, weeks, months to react to a new vulnerability, because human attackers are relatively slow) can collapse when a model helps reverse-engineer a patch and weaponise it at speed. The fundamentals don’t matter less in that world. They matter more, and faster. If anything, Glenn and Sam agree, that’s an argument for more skilled defenders, not fewer — someone still has to patch the box and keep the inventory honest. We’ve seen this rush before Glenn has been in the game long enough to recognise the pattern. Dot-com. Cloud. IoT. Web3 and crypto. Each time, the same story: everyone rushes to adopt the shiny new thing, and security is a problem for later. AI is simply the current chapter. Fools rush in where angels fear to tread — and hackers love it, because every rushed rollout means a bigger attack surface and more complexity to hide in. The Langflow box is the archetype. Somewhere, someone was told “we need to do AI,” stood up a server, left the default credentials in place, skipped the updates, and moved on. Quite possibly shadow IT the sysadmins didn’t even know was running. Layer upon layer of ordinary risk, created by the pressure to move fast. Tools, not jobs For all that, Glenn is not a doom-monger — and he’s pointed about the anxiety the hype is causing. He gets graduates coming up after talks asking whether they should abandon cybersecurity for a plumbing diploma. His answer, drawn from watching tools come and go for decades (he remembers the same end-of-the-world panic when Metasploit first landed): AI will replace tools, not people. “Be the AI wrangler. Maybe in five years you manage fifty agents — but you’re the one in charge.” Glenn Wilkinson, on advice for the next generation And a neat economic backstop for the anxious: if all the humans are replaced, who’s putting cash into the economy to buy what’s being built? He’s just as sharp on why the guardrails around these models keep failing. Prompt engineering is now a genuine skill — and so is prompt injection. The reason jailbreaks are such an endless game of whack-a-mole, he explains, is structural: the industry largely tamed SQL injection by learning to separate data from commands. Large language models put your instructions and your data in the very same channel. Until that’s solved, tricking a model into confusing the two will remain frustratingly doable.PODCAST: The Resilience Room
https://www.immersivelabs.com/resources/the-resilience-room 👆 Listen on your favourite podcast app or online here Welcome to The Resilience Room, where cyber professionals sit back and chat about their lives, passions and experiences. We discuss cyber culture, thought leadership, technical topics and emerging trends. Hosted by Sam Dickison, Community Manager at Immersive. 💡 We'd love to hear your questions for guests, or guest suggestions! Please comment on this post with any ideas.
