Briefing
Episode 29 · Primer and discussion · About 14 minutes in two parts
Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.
Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.
Written from CrowdStrike’s blog and report page and press coverage of the full report, which we could not read, checked on 5 November 2026. This is one security vendor’s threat intelligence, and the vendor sells defences against this threat. Its growth figures are not used. Check the original before quoting.
Before we record, every claim the hosts make is checked against the sources (this table was last updated 5 November 2026). Confirmed means a primary source or two independent reports; Reported means one party says so; Partly right, Disputed, Unverified and Not yet known mean what they say, and the hosts say so on air. This is the table as we left it.
Sign up for our podcast at agile-insights.com.au.
Linda Here is the idea. Attackers are now running whole break-ins with the same A I coding agents that organisations are adopting. They set them up the way we set up our own, with a playbook, a kit of skills and broad permission. So the controls that matter are the ones that still hold when an agent moves in minutes. By the end of this part, you will know what an agent and its harness are, what CrowdStrike found, and why the model's safety rules are not your control.
Leonard This comes from research published on the twenty first of October by CrowdStrike, the security company. Its threat intelligence team described four groups it had not reported on before. Between May and August this year, each one used A I agents to make decisions and run commands inside real intrusions. Before we get to what they did, I need the basics. When people say agent, what do they actually mean?
Linda An agent is a language model given three things: tools, a goal and a loop. The tools let it act, for example by running commands on a computer. The goal says what done looks like. The loop lets it look at the result of each step and choose the next one itself. A chatbot answers you. An agent keeps going. A plain example is a coding agent fixing a bug. It reads the code, runs the tests, sees the failure, changes a file and runs them again, without asking you each time. All four groups in this report used Claude Code, Anthropic's coding agent, which many developers use every day.
Leonard It keeps going. Right.
Leonard So the model is the engine, but something has to keep it on the road for hours at a time. The report keeps coming back to one word for that, which is harness. What is a harness?
Linda A harness is everything wrapped around the model to keep it on task. That means instructions, a playbook, saved progress, and a library of skills, which are short how-to files the agent can pick up. The first ran two agents. A supervisor picked the next stage of the attack from a fixed workflow, and an executor carried it out. The fourth group used one command that handed the agent a seven step plan, with near total freedom over how to break in. The third, which CrowdStrike calls the least sophisticated, downloaded eight hundred and nineteen skills, mostly from public code sites, and told the agent to get in...
Leonard That third group is the one that worries me. According to CrowdStrike, it still compromised several government web domains in Indonesia and another country, plus private companies. So building a capable attacker is no longer the hard part. CrowdStrike says the main limit on scaling is not engineering skill but token budgets, meaning the bill for running the model. But surely the model refuses to do this? Isn't that what safety training is for?
Linda It is meant to, and that is the third idea. Guardrails are the model maker's rules for refusing harmful work. CrowdStrike reports that all four groups got past them. According to press coverage of the full report, the usual trick was simple. The operators said the work was an authorised penetration test, which is a security check a company pays for. Some left files in the agent's working folder saying every target was authorised. One used a modified version of the tool to strip out its refusal instructions. Anthropic described the same pretext in a case it disclosed last November, where attackers told Claude they worked for a legitimate security firm.
Leonard And then there is the speed. CrowdStrike's own headline finding is about one agent inside a victim's network. It moved across that network and harvested the credentials for the domain controller in about twenty minutes, with no further human help. For anyone new to this, a domain controller is the server that holds the keys to every account in an organisation. CrowdStrike says these agents often reach their objective within an hour.
Linda Twenty minutes.
Linda Here is the common misunderstanding. People hear this and conclude the A I is a genius hacker. An agent tries every door quickly and patiently, and the door it gets through is usually one we left open. In one case from press reports of the full report, an agent found a cloud access token that carried administrator rights over a whole cluster of servers. It used that to reach hundreds of databases. That is an over privileged identity, not a clever trick. So the damage comes from two things we control: how much each identity can do, and how fast we notice.
Leonard One caution before we go on, because listeners will be thinking it. This is one security company's research, and that company sells defences against exactly this threat. So how much of it do we actually know, and how much is still the vendor's view?
Linda Here is what is confirmed, and what is not. CrowdStrike's own pages state the four groups, how they set up their agents, the twenty minute case, and that all four bypassed the safety controls. Some details, like the hundreds of databases, come from press reporting of the full report. CrowdStrike saw inside because the attackers left their folders exposed online. What nobody has said is how many organisations were hit in total, or whether any victim caught it themselves. And its forecast, that these attacks will almost certainly grow, is an assessment, not a measurement.
Leonard That is a fair line to hold. Which brings me to the question for everyone listening who has to brief a board. What does this mean for governing A I inside your own organisation?
Linda Three things. First, model guardrails belong to the supplier, and an attacker can talk past them, so do not count them as your control. Treat anything an agent reads as untrusted, including a file in its own folder. Second, give every identity, whether human, service or agent, only the rights it needs, and check the tokens, because that is what an agent will find. Third, measure detection in minutes, not days. A daily review is too slow for an attacker that finishes within the hour.
Leonard That is the primer. In Part 2, we argue about what actually changes, including the agents your own staff already run.
Linda Here is the idea again, for anyone joining us now. Attackers are running whole break-ins with the same coding agents organisations are adopting, and they set them up the way we set up our own. So the controls that matter are the ones that hold at machine speed. In this part we argue about what that changes. Is this ordinary security done faster, or does it change how an organisation governs its own agents? By the end, you will have a short list to take to your security and A I leads.
Leonard First, a thirty second recap. An agent is a model with tools, a goal and a loop, so it chooses its own next step. A harness is the playbook and skills wrapped around it. On the twenty first of October, CrowdStrike reported that four groups used Claude Code harnesses in real intrusions between May and August. All four talked past the model's safety rules, mostly by claiming the work was an authorised test. And one agent reached the keys to a whole network in about twenty minutes, without further human help.
Linda So the question is whether any of this is really new. Leonard, I think the sceptic's answer deserves to go first, so make the case.
Leonard Happy to. Look at what the agents actually exploited. A stolen password, reused. A cloud token with far too many rights. Flaws in web applications. None of that is new. Security testers have used automated tools for decades, and attackers have copied them for just as long. On that reading, the fix is the security basics we already know, done properly. Patch, tidy up permissions, watch the logs. A I governance, on this view, has little to add, and calling it an A I problem lets ordinary security debt off the hook...
Linda I agree with most of that, and I think it misses the change that matters to us. The techniques are old. Who can run them, and how, is new. The least skilled group in the report still broke into government domains, using downloaded skills and a model. CrowdStrike assesses, with moderate confidence, that agentic A I commoditises baseline offensive operations. And look at how they set up their agents. A supervisor and an executor. A standard operating procedure. A skills library. A seven step command. That is an operating model for agents. The attackers wrote theirs in a few months. Many organisations are still drafting their own.
Leonard That is fair, and it is the part I find uncomfortable. Our own developers write exactly that kind of command for the same tool. So the governance question is not only how we defend against someone else's agent. It is whether our own agents could be pointed the same way. One detail in the reporting shows why. The operators told the agent its targets were authorised, sometimes with a file left in its own working folder. An agent inside a company reads files, emails and tickets all day. If a planted line can grant permission, then everything an agent reads is a possible instruction.
Linda Which is why I would separate two kinds of control. Guardrails are the model maker's refusals. They matter, and the operators in this report still had to work to get round them. CrowdStrike says the least skilled group had to step in from time to time to push the agent past a refusal. But guardrails are someone else's control, and an attacker can argue with them. Your controls have to live in your own estate. What an identity is allowed to touch. What an agent treats as an instruction. How fast someone notices when it does something new. Those hold whether or not the model says no.
Leonard Someone else's control.
Leonard Here is the inside half of the story. At its conference in September, CrowdStrike's president said one large customer had found eighteen thousand A I agents on its computers, against three hundred it had approved. That is a vendor's anecdote about one unnamed company, and we could not check it. But the question it raises is fair. Do you know how many agents your staff run, under whose identity, and with what access? If not, the over privileged token that an attacker's agent finds might belong to one of yours.
Linda Quite a gap.
Linda So where does the evidence settle? Here is what is confirmed, and what is still open. CrowdStrike's own pages establish the four set ups, the twenty minute case and the bypassed safety controls. Anthropic disclosed a similar campaign last November, in which it said A I did eighty to ninety per cent of the work. What is not settled is scale. We have one vendor's view, seen because the attackers left their folders open, from a company that sells defences against this. CrowdStrike reports rapid growth, but it has not published how it counts. So act on the case studies, and do not quote the growth numbers.
Leonard Where I still differ from you a little is on speed. Twenty minutes sounds dramatic, and it makes a good headline. But I suspect the bigger gap for most organisations is not speed. It is that nobody looks for weeks. If that is the real gap, a stopwatch is the wrong thing to show a board. The fix is still the dull discipline of actually reviewing what happened...
Linda Both are true, and they point the same way. If no one looks for weeks, speed hardly matters, so start by making sure someone looks. But once you do look, a daily review is too slow for an attacker that, in CrowdStrike's words, often reaches its objective within an hour. So the target is detection in minutes for the identities that can do the most damage, starting with agents and service accounts. And you need a tested way to stop an agent quickly, not a plan to write one after the first incident.
Leonard Agreed. Someone has to look, and then they have to look fast.
Linda Here is the list to take away. One, count your agents, give each one an owner, and know whose identity each one uses. Two, cut every identity, human, service or agent, down to the rights it needs, and hunt for tokens with administrator power. Three, treat everything an agent reads as untrusted, including files in its own folder, so no planted line can grant permission. Four, do not record the model's safety rules as your control. Five, measure detection in minutes for your most powerful identities, and test that you can stop an agent fast. And when you brief your board, quote the cases, not the growth figures.
Leonard The attackers have already written their playbook for agents, in a few months, with tools anyone can download. The useful response is to write ours, and to make sure it holds at their speed, not ours.
Linda If you'd like to know more, sign up for our podcast at agile-insights.com.au.
Leonard Thanks for listening.
Every episode tagged #security
All of the briefings (38) · Follow Briefing in Apple Podcasts or any podcast app
Everything from Agile Insights: briefings, Making You Think and the reference series