Inside the AI-Driven Insider Threat

Simon Carver

[warmly] Welcome to the show. A company can spend millions hardening its perimeter, and still get burned by the person who already has a badge, a laptop, and now an AI copilot. That’s the quick take today: the insider threat has gone 10x faster. I’m Simon Carver. If this kind of conversation helps you think more clearly about AI and work, like, share, and subscribe. And joining me for this one is Lachlan Reed.

Lachlan Reed

[curious] G’day. And yeah... this one’s a bit spicy, isn’t it? Because for years, the mental picture was some hoodie-wearing hacker outside the fence. But the fence is not the whole game anymore. The tricky bit now is the person inside the yard, using the same AI tools the company rolled out for “productivity gains.” [chuckles] Bit of a snake in the toolbox situation.

Simon Carver

[reflective] That phrase from the outline sticks with me: the most dangerous threat may already be inside the company. Not because every employee is a villain. Most aren’t. But because the old assumption was external threat first, internal threat second. And AI flips the leverage. A person with legitimate access can now map weak controls, test where audits are thin, and move much faster than managers expect.

Lachlan Reed

[matter-of-fact] Faster and quieter. That’s the bit leaders miss. This isn’t just old-school expense fraud with fancier window dressing. It’s industrialized internal fraud, or data abuse, juiced up by automation. Same keyboard, same employee login, same normal-looking day on the calendar... but under the hood, the pace is completely different.

Simon Carver

[questioning tone] And the source material makes a pretty blunt point: the same tools meant to drive efficiency can also accelerate fraud. That’s the tension, right? You hand someone a system to summarize documents, draft code, search across knowledge, maybe even connect workflows. Those are useful powers. But useful powers are also exploitable powers.

Lachlan Reed

[chuckles] Yeah, a torque wrench can fix the bike or strip the bolt if you go too hard. Same tool. Different intent. And I reckon what makes this uncomfortable is that leaders still talk like risk lives at the perimeter -- firewalls, phishing, outsiders knocking on the door. Meanwhile the real drama may be happening in a perfectly ordinary employee account that looks, at a glance, boring as bathwater.

Simon Carver

[softly] Which is why this conversation can’t be alarmist, but it also can’t be naive. If your model still starts with “we trust insiders by default and check occasionally,” you may already be defending the wrong edge.

Lachlan Reed

[excited] Right, so let’s name the thing: the agentic insider. That’s an employee -- or even a small network -- using AI agents to think faster, test faster, and exploit faster than legacy controls can keep up with. Not just one clever bloke poking around after hours. More like having a tireless digital offsider running hundreds of little experiments while the company is still waiting for next month’s audit.

Simon Carver

[skeptical] “Next month’s audit” is the part that lands for me. Because most controls really are built around delay. Access controls are static. Monitoring is reactive. Audits are periodic. Static, reactive, periodic -- those are three very old words in a world where AI systems are continuous.

Lachlan Reed

[responds quickly] Exactly. Continuous is the key word. The insider doesn’t have to smash through a wall anymore. They can study the wall. Study the camera. Study what gets flagged. They use AI to learn alert thresholds, blind spots, timing patterns. Then they operate just under the line. Not loud enough to trigger the obvious alarm, just steady enough to blend in.

Simon Carver

[leaning in] “Just under the line” -- that is such a nasty detail. Because a lot of organizations still look for legacy signals: unusual login times, large file downloads, access outside normal patterns. But if I’m hearing you right, the modern insider doesn’t merely avoid those signals. They model them.

Lachlan Reed

[matter-of-fact] Spot on. They don’t trip alarms -- they study alarms. Let me try a concrete version. Say the system tends to notice one huge download. Fine. So instead of one giant grab, activity gets broken into normal-looking pieces. Or if the system watches for weird timing, the work gets pushed into normal business rhythms. Same goal, cleaner footprint.

Simon Carver

[pauses] And there’s an even darker twist in the source: creating noise. Not just hiding under thresholds, but flooding detection with junk patterns so the real signal gets buried. That’s not just evasion. That’s manipulation of the detector itself.

Lachlan Reed

[grimly] Yep. Simulated noise. False alerts, distractions, lots of harmless-looking activity. Like kicking up dust at the far end of the paddock so nobody notices the gate open at the back. And because AI can generate that noise cheaply, the defender ends up chasing ghosts while the real activity strolls through in work boots.

Simon Carver

[serious] The examples here are not theoretical-sounding either. Synthetic identities entering organizations with AI-generated histories. Ghost employees collecting paychecks with no real presence. Vendor collusion hidden through relationship complexity. Data exfiltration through seemingly harmless AI prompts. Each one of those rides on something organizations already do: hiring, payroll, vendor management, internal tools.

Lachlan Reed

[hesitates] The “ghost employees” one gets me. Because that sounds almost too old-fashioned to still work -- like some dodgy scheme from a bad movie -- but pair it with AI-generated histories and suddenly the paperwork can look polished. Clean LinkedIn-style story, neat chronology, all the little bits lined up. Even a kangaroo could trip over that if the checks are lazy.

Simon Carver

[thoughtful] And this is where I want to push a little. Is the real novelty AI itself, or is AI just exposing how shallow some controls were all along?

Lachlan Reed

[short pause] I’d say the second one is more important. AI is the accelerator, not the original sin. The original sin is pretending periodic review was enough. If your oversight only wakes up every quarter, then 10x speed doesn’t just make things faster -- it changes the game board. The gap becomes the strategy.

Simon Carver

[calm] That brings us to the real diagnosis: this is a governance failure. Organizations raced toward “10x productivity” without redesigning control frameworks, risk models, or accountability structures. So now they have 10x capability sitting on top of 1x oversight. That ratio -- 10 to 1 -- is the memorable one for me.

Lachlan Reed

[firm] Same here. Ten-to-one is where the trouble lives. And it matters because people love to say, “Well, this is a technology problem.” Nah. Tech exposed it, sure. But governance set the table. If leadership pours high-octane AI into the engine and leaves the brakes stock, don’t act shocked when the ute ends up in a ditch.

Simon Carver

[reflective] Here’s the hard part, though. The countermove sounds like surveillance if you say it clumsily. Behavior-based monitoring. Continuous control validation. Relationship mapping. Real-time intervention. Those phrases can make employees feel like the office has turned into a panopticon. And if leaders handle that badly, they damage the very trust they claim to protect.

Lachlan Reed

[skeptical] True -- and we shouldn’t brush that off. Nobody wants to work in a digital fishbowl. I’d hate it. But the choice isn’t “trust everyone” or “spy on everyone.” That’s a furphy. The real choice is between assumed trust and verified behavior. Between titles and tenure on one side, and patterns and evidence on the other.

Simon Carver

[questioning tone] So let me say it back, and you tell me if I’m oversimplifying. We’re not moving from trust to suspicion. We’re moving from trust as a feeling to trust as a system. Is that close?

Lachlan Reed

[warmly] Yeah, that’s close. Trust as a system is good. Because if a small number of bad actors now have amplified impact, ignoring that doesn’t preserve culture -- it endangers culture. Good employees get hurt too when controls fail. Payroll fraud, bad vendor deals, data misuse... that all lands on real people eventually.

Simon Carver

[softly] That’s the human cost that gets missed. When governance is weak, the damage doesn’t stay abstract. It shows up in layoffs, reputation hits, broken teams, and a creeping sense that leadership was asleep at the wheel. So the practical answer has to protect both the business and the workforce.

Lachlan Reed

[matter-of-fact] Which means a few things. Behavior-based monitoring -- what the source calls seeing patterns, not just actions. Continuous control validation, so you’re not waiting for an after-the-fact audit to tell you what already happened. Graph-based relationship mapping, so collusion and hidden networks don’t just look like random dots. And real-time intervention, because last quarter’s report is about as useful as yesterday’s weather when the fraud is happening now.

Simon Carver

[warmly] And maybe the biggest requirement of all: leadership maturity. Leaders have to explain why these systems exist, what they watch, what they do not watch, who is accountable, and where the guardrails are. Otherwise “protecting integrity” becomes a blank check, and blank checks are dangerous in any direction.

Lachlan Reed

[reflective] So maybe that’s the line to leave people with: in the AI era, the workforce is human plus machine augmentation. Governance has to match that reality. If it doesn’t, you’re not managing the future of work -- you’re just hoping it behaves itself. And hope’s not a strategy, mate.

Simon Carver

[warmly] If you liked this quick take, subscribe and share it with someone thinking seriously about AI, work, and risk. We’ll see you next time.