Latest / Decrypted: The UK Cyber Briefing / PHP2Shell and Node2Shell: the day the web's engines both broke
Transcript
- 0:00Annabel: This is Decrypted. Sunday the twenty-fifth of October. I'm Annabel.
- 0:05Jude: And I'm Jude.
- 0:06Annabel: And this is a special edition, because this week-
- 0:10Jude: This week was mad.
- 0:12Annabel: This week was mad. Two of the engines that run a huge part of the web both had emergency fixes on the same Wednesday. PHP, and Node.
- 0:23Jude: And both of them got the Shell treatment. PHP2Shell. Node2Shell.
- 0:29Annabel: Do you want to explain the Shell thing? Because you were there for the first one.
- 0:33Jude: I was, yeah. December twenty twenty-one, Log4Shell. A bug in a Java logging library. If you could get one line of text written into a log, which is nothing, you could take over the server. It was everywhere. I spent most of that Christmas on the phone. And ever since, if a bug ends in "Shell", it means an attacker can get a shell on your machine.
- 0:56Annabel: And a shell, for anyone who's never-
- 0:58Jude: Command line. The place you type instructions. If someone else has a shell on your server, it's their server now. You're just paying the bill.
- 1:06Annabel: You're just paying the bill. Right. So before we go anywhere, here's the honest headline, because I think the coverage this week lost it. Both bugs are serious. Both are fixed. One of them was being attacked before the fix even existed, and it took down bits of the British state as a precaution. The other looks a lot scarier on paper than it has turned out to be, so far. And both of them, and this is the bit I can't stop thinking about, were found by AI.
- 1:38Jude: Both of them.
- 1:40Annabel: Both. We'll get there.
- 1:41Jude: Can I start with what I saw, rather than what the bug is? Because I think that's how most people met this.
- 1:47Annabel: Go on.
- 1:47Jude: Wednesday morning. I'm trying to report something on the Met Police website and it's just gone. Then my mate who works for a council texts me, their site's down. Then it's on the news. Edinburgh, Oxford, Bristol, Lambeth, Barnet. The Office for Budget Responsibility, which is the Treasury's independent forecasters, their website's off. BBC portals. NHS Scotland. Six of the fourteen Scottish health boards.
- 2:14Annabel: Seven Scottish councils.
- 2:16Jude: Seven Scottish councils. And Edinburgh described it as "essential maintenance".
- 2:21Annabel: Everybody's essential maintenance happened to fall on the same morning.
- 2:26Jude: Funny that. So what was the actual scale? Because at the time nobody would say.
- 2:32Annabel: Thursday afternoon, the House of Lords, an urgent question. The minister, Baroness Lloyd, said thirty-two central government services were unavailable and four partially. She said the NCSC, that's the National Cyber Security Centre, the public-facing bit of GCHQ, had sent fifty thousand proactive warnings across the economy. And they weren't aware of any confirmed compromise of government systems. The impact, in her words, was almost all from precautionary measures.
- 3:06Jude: So the state switched itself off on purpose.
- 3:09Annabel: Bits of it, yes. And the NCSC's own alert on the Wednesday literally said "consider taking the system offline" if you couldn't patch or mitigate.
- 3:20Jude: See, that's where I'm not sure. The Dutch did the same, their national cyber centre and one of their big tech sites went offline, and I remember thinking, that's a bit dramatic. And then we did it too. Was that the right call, or was that panic?
- 3:35Annabel: I think it was the right call and I'll tell you why in a bit, but I want you to hold that question, because it's a better question than it sounds.
- 3:43Jude: Noted.
- 3:44Annabel: So. PHP. If you've never written code you've still used it today. It's the language behind WordPress, most online shops, forums, webmail, wikis. On the counting sites it's behind roughly seven in ten websites where you can tell what's under the bonnet.
- 4:02Jude: Which means when PHP itself breaks-
- 4:05Annabel: Every site built on it has the problem, however carefully that site was written. That's the whole story in one line. These aren't bugs in somebody's plugin. They're in the engine. And they go back to PHP five point four. Twenty twelve.
- 4:22Jude: Fourteen years.
- 4:24Annabel: Fourteen years. Two flaws, scored nine point two out of ten on the CVSS scale, which is the industry's zero-to-ten severity score.
- 4:34Jude: Okay. The bugs. I think I understood one of them and I want to check, because I've been reading.
- 4:40Annabel: Go on then.
- 4:41Jude: So PHP doesn't start fresh for each visitor. One worker process handles request after request. And there's a sort of sticky note inside it that says "the cash box is in drawer three". Drawer three gets emptied and handed to someone else, but nobody takes the note down-
- 4:59Annabel: And the next request reads the note.
- 5:01Jude: And trusts whatever's in drawer three now.
- 5:04Annabel: Which the attacker has filled. Yeah. That's it. That's genuinely better than how I'd have put it.
- 5:10Jude: I've been reading Sansec.
- 5:12Annabel: You've been reading Sansec, who are the people who watch online shops for a living. So, that's the stale state bug. One confirmed way to leave the note behind is a feature called upload progress, which shows you a progress bar when you upload a file. It's on by default. And it does its work before your website's own code runs, so a site with no upload button anywhere is still in.
- 5:38Jude: Which I did not know until Wednesday.
- 5:40Annabel: Nobody did. That's rather the point. The other bug is the out-of-memory one. Every PHP process has a memory ceiling. The attacker sends a request built to smash through that ceiling halfway through PHP setting up, and PHP bails out so fast it skips the tidying. Half-built things left lying around in memory, and the attacker's data sitting where PHP expects something it can trust.
- 6:07Jude: And it runs it.
- 6:09Annabel: And it can be steered into running it. Those are the two. Now, the thing people got wrong this week was mixing them up, because the tool that leaked folded them together. That matters because the stopgaps are different. Switching off upload progress helps with the first one and does nothing for the second.
- 6:28Jude: Right, the leak. Because I don't think people realise how this was supposed to go.
- 6:33Annabel: How it was supposed to go: researchers find it in August, report it privately, PHP writes the fix quietly, announces a date, everyone patches, then the details come out. That's coordinated disclosure, that's the grown-up way. And it nearly worked. On Tuesday the twentieth PHP announced an emergency release for two o'clock Wednesday afternoon. Same day, the details leaked, with a working tool. PHP moved its full disclosure forward a day.
- 7:04Jude: And the attackers were already moving.
- 7:07Annabel: Attackers were already moving before the leak. Amazon's honeypots, the decoy servers they run to watch attackers, saw an attempt on the seventeenth. Four days before anyone public knew.
- 7:19Jude: Four days before.
- 7:21Annabel: Then on the Wednesday, the threat intelligence firm ELLIO saw the first scanning hit their sensors just after half past eight in the morning. Five and a half hours before the fix was out. Cloudflare, who sit in front of a huge slice of the web, had blocking rules live at two minutes past three that morning, nine hours before PHP published. By eleven that night ELLIO had counted seven hundred and forty-three attack requests from thirty-three addresses, and by Thursday morning some were carrying real malware.
- 7:54Jude: Okay, so, going back to my "was it panic" question-
- 7:58Annabel: You held it.
- 8:00Jude: I held it. If attacks were landing five hours before a fix existed, then for five hours there was no patch to apply. You could either be off, or be a target.
- 8:11Annabel: That's exactly it. And for most public bodies there's a second problem. They don't run their own PHP. Their website lives with a hosting company or an agency. They can't patch it themselves. So for them "wait for the host" or "switch it off" were the only two buttons. The Dutch pressed off. So did Edinburgh.
- 8:33Jude: Fine. I'll give them that one.
- 8:36Annabel: I'm going to remember you said that.
- 8:38Jude: Can I do the chief executive question? Because I think everyone listening who runs something, not a website, a company, is thinking "fine, websites and forms, I've got a web team for that". Where else is this?
- 8:50Annabel: Good. Let's walk in from the outside. The public website, obviously. Then the things behind it you forget are websites. The hosting control panel, cPanel or Plesk, is written in PHP. Webmail. The tool your IT team uses to poke the database. Your wiki, your ticketing system, the monitoring dashboard on the wall.
- 9:13Jude: The intranet.
- 9:14Annabel: The intranet. Then appliances, boxes you don't think of as computers. The web page you log into on your firewall, your network storage box, your phone system. Some of those are PHP. One open-source firewall, OPNsense, patched on day one, and on that product PHP runs as root, the all-powerful account. The security firm Sophos said none of their products was affected and then took one of their own old customer portals offline because it ran PHP.
- 9:47Jude: Everyone's got one.
- 9:49Annabel: Everyone's got one. Now, the question you asked me on Wednesday-
- 9:54Jude: Linux. Because the headlines basically said "Linux is vulnerable" and I've got Linux boxes with no websites on them.
- 10:01Annabel: So, no. Linux the operating system is not vulnerable. PHP is a package you install on it. A box with no PHP on it is out of scope, full stop. But, and this is the bit for the CEO, the big Linux distributions all ship PHP, and they patched at different speeds. Debian and Ubuntu on the day. Red Hat said some updates were available, more coming. So "we're on Linux" tells you nothing. "Which packages, which version, has it restarted" is the question.
- 10:37Jude: And old stuff.
- 10:38Annabel: Old stuff is the uncomfortable number. PHP only patches the versions it still supports, which is eight point two and up. Eight point one and older, all of PHP seven, all of PHP five, nothing from PHP. One scanning firm, CybelAngel, reckons fifty-nine percent of the PHP hosts they can see are on versions that will never get a fix. By the counting sites, about a third of all PHP websites.
- 11:08Jude: A third of the web's PHP sites can't be patched.
- 11:14Annabel: Not from upstream. Some Linux distributions and paid support firms are back-porting fixes. Otherwise it's migrate or be exposed. And for scale: Censys, who scan the whole internet, counted about thirty-six point seven million internet-facing services on affected PHP versions on the Thursday. Another firm found ninety-nine point three percent of what they could measure still unpatched a day in.
- 11:41Jude: Ninety-nine point three.
- 11:43Annabel: It'll be better now. It won't be good.
- 11:46Jude: Last one on the map. Internal systems. The old customer database nobody outside can see.
- 11:52Annabel: In scope, lower priority. The NCSC's words were "prioritise internet-facing systems first and then internal systems". Not "ignore internal". Because the thing a CEO needs to hear is: visible and reachable aren't the same. A phished laptop, a supplier's remote link, a misrouted connection, and your invisible intranet is reachable. Both of these bugs need nothing but a request arriving. No login.
- 12:21Jude: Let's do the public sector properly, because that's where the week actually happened. What does government run on?
- 12:28Annabel: Public sources only, and "runs on" is not "is vulnerable", I want to be careful.
- 12:34Jude: Understood.
- 12:35Annabel: GOV.UK itself, the publishing platform, is mostly Ruby on Rails. Not PHP. But the common bits that departments build with lean on Node. The GOV.UK Frontend toolkit is a Node package. GOV.UK Pay, the payments service, has Node front ends. One Login, the new sign-in for government, is partly Node. HMRC is on something else again. So central government's own platforms were more a Node week than a PHP week.
- 13:05Jude: Then why were thirty-two services off?
- 13:08Annabel: Because departments have hundreds of smaller sites, arm's-length bodies, campaign sites, consultations, and an awful lot of those are WordPress or Drupal, both PHP, often with an agency, often on shared hosting. The OBR site is a good example. Nobody published the list of thirty-two.
- 13:29Jude: Councils.
- 13:30Annabel: Councils are the PHP estate. Drupal is now the most common council website platform in the UK and Ireland, around a hundred of four hundred and ten councils, and there's WordPress on top of that. So something like three in ten council websites are PHP. Which is why the names you read out on Wednesday were councils.
- 13:52Jude: And the Met? Because that's the one that got me.
- 13:55Annabel: This is interesting. Forty of the forty-three police forces in England and Wales, including the Met, use a shared platform called Single Online Home, and it's publicly described as running on a Microsoft-based system, not PHP, behind Cloudflare. So why the Met's site went off, nobody has said publicly. It might have been a different Met property, it might have been a platform-wide precaution. I don't know, and nobody's told us.
- 14:24Jude: So I'm allowed to be slightly annoyed about that one.
- 14:27Annabel: You're allowed to want an explanation, yes. NHS: NHS England's national security operations centre put out a high-severity alert on the Wednesday afternoon, rating future exploitation "highly likely". A lot of NHS trust websites are WordPress. And Scotland, as you said, six health boards went dark as a precaution.
- 14:52Jude: Defence?
- 14:53Annabel: Nothing public. No Ministry of Defence site reported offline, no statement I could find. Where I'd worry in defence is the supply chain, the thousands of small suppliers whose portals sit on exactly the shared hosting we've been talking about.
- 15:09Jude: And money. Because the Met site being down is annoying. A bank being down is a different kind of day.
- 15:16Annabel: The banks that talk about their technology publicly describe Java cores, not Node or PHP. And none of the regulators, the Financial Conduct Authority, the Prudential Regulation Authority, the Bank of England, said a word about either bug this week. The one hard deadline that applies is the card industry's rules, PCI DSS, which say critical patches go on within a month. So anyone taking card payments through PHP has until about the twenty-first of November.
- 15:47Jude: A month. For a bug being attacked five hours before the patch.
- 15:52Annabel: A month is the outer bound. It's not advice.
- 15:56Jude: Right, so, big picture, how prepared was the state for this?
- 16:00Annabel: Honestly? The National Audit Office looked at government cyber resilience in January last year. The programme that assesses critical government systems, GovAssure, had covered fifty-eight of them and found low maturity in the basics. Knowing what you've got. Watching it. Planning for when it breaks. And two hundred and twenty-eight legacy systems weren't in the programme at all.
- 16:25Jude: Knowing what you've got.
- 16:26Annabel: Knowing what you've got. Hold that one too, because it's the answer to the whole episode.
- 16:32Jude: Node. Same thing in a different language?
- 16:35Annabel: Different language, different bug, and a very different week. Node is how a lot of modern web applications run on the server, using JavaScript, the browser's language, on the back end. And the flaw isn't even in Node's own code. It's in something called Brotli.
- 16:52Jude: Which is a vegetable.
- 16:54Annabel: It is not a vegetable. Brotli is a compression format from Google. Web pages get squashed with it to save time and unpacked at your end. Vacuum-packing a duvet to get it in the post. Node carries its own copy of Google's unpacker, and the unpacker had a bug. A specially built package tells it to switch between kinds of data in a way it shouldn't allow, the old code notices the error and then carries on anyway, writing past the end of its buffer into memory it doesn't own.
- 17:27Jude: And writing where you shouldn't is how you go from "crash" to "run my code".
- 17:32Annabel: In the right hands. Node pre-announced on the fourteenth of October and shipped the fixes on the twenty-first, a few hours before PHP's. Twenty-two point twenty-three point four, twenty-four point twenty-two point one, twenty-six point twelve point one. Bun, another JavaScript runtime, fixed it in one point four point four.
- 17:54Jude: Who's exposed? Because not every Node app unpacks Brotli.
- 17:58Annabel: Two main routes. Express, which is the framework most Node websites are built on, so if you've got a Node app it's probably in there. Version five unpacks compressed requests by default. Version four doesn't. And the sneaky one: when a Node server fetches something from another website, Node unpacks a Brotli reply if the reply says it's Brotli, even if your code asked for it plain. So if your server fetches links users give it, a hostile site can hand you a booby-trapped reply.
- 18:31Jude: Okay, now the bit where we disagree.
- 18:34Annabel: Do we?
- 18:35Jude: I think we do. Node rated this "high". Google's record scores it nine point two, critical. And there are researchers saying, realistically, it's a crash-your-server bug, and taking the server over is much harder than the headlines. I'm with the researchers. I think this got the Shell name because PHP got it the same day.
- 18:55Annabel: And I think you're half right, which is annoying. Here's the evidence. Imperva, a security firm, got code execution on an official Node image, one version, one chip type, and it took them several attempts. The public proofs of concept crash servers reliably. One team said blind remote execution with normal memory protections on didn't work for them. No confirmed exploitation in the wild as of Friday. Not on the American list of known exploited bugs.
- 19:26Jude: So I'm right.
- 19:28Annabel: You're right that it's not PHP2Shell. You're wrong that it doesn't matter. The security firm Qualys counted three point three million Node installations they could see. One point three million on supported versions, unpatched. One point seven million on end-of-life versions that will never get a fix. Forty-two thousand patched.
- 19:50Jude: Forty-two thousand out of three million.
- 19:54Annabel: And "an attacker can knock over your checkout at will" is not a low-severity sentence, Jude.
- 20:00Jude: No. Fine. Call it a draw.
- 20:05Annabel: It's not a draw. Two more Node things. One, Node is in places you forget. Build servers, serverless functions, and every desktop app built on Electron, so VS Code, Slack, Teams, Discord, carries a copy. Whether this bug is reachable in those apps is genuinely open. Nobody had said by this weekend.
- 20:28Jude: And two?
- 20:29Annabel: Two. Nobody in the UK said anything about Node2Shell at all. No NCSC alert. No statement. The Dutch national cyber centre put one out. So did the Belgians. Ours didn't.
- 20:43Jude: Huh.
- 20:44Annabel: Hold that one as well.
- 20:45Jude: I'm holding a lot of things.
- 20:47Annabel: Okay. The bit I said I couldn't stop thinking about. Who found these bugs?
- 20:53Jude: Researchers.
- 20:54Annabel: Researchers with AI. A small firm called Calif found the first PHP bug on the first of August using one of OpenAI's models. Wiz, a big cloud security company, found it independently on the nineteenth of August using several frontier models, Google's and OpenAI's and their own. Reported it on the twenty-fourth. PHP acknowledged it in September and patched it on the twenty-first of October.
- 21:23Jude: Hang on. Two months.
- 21:25Annabel: Two months from report to fix. And Node2Shell was found by OpenAI's own Codex Security team. Both of this week's bugs. Code that had sat in PHP for fourteen years with humans reading it, found inside a fortnight once you pointed a model at it.
- 21:44Jude: And this is the thing the minister mentioned, isn't it? Because I remember that line.
- 21:48Annabel: She said the NCSC "was briefed by OpenAI last week". Before it was public. And the AI minister, Kanishka Narayan, said an earlier AI-assisted review of government systems had found five hundred vulnerabilities.
- 22:04Jude: Five hundred. In government systems. Found by a chatbot.
- 22:09Annabel: Found by a model, yes. And Richard Horne, who runs the NCSC, had a line about this back in April. He said these models "are not finding new attacks, they're just exposing more security vulnerabilities". Which sounds like a dismissal and isn't. It's a warning. The bugs were always there. We just couldn't see them.
- 22:33Jude: Does the finding actually hold up, though? Because last year the story was AI "slop", wasn't it. Open-source maintainers drowning in rubbish AI bug reports.
- 22:44Annabel: It was, and this is the bit that's changed. Daniel Stenberg, who maintains curl, a tool that's in basically every computer on earth, was the loudest voice on slop. Last year under five percent of the AI reports he got were real. This Wednesday he wrote that since the spring the slop is "practically gone". Every report he gets now has AI in it somewhere, and they're, his words, "typically accurate and correct, even if much too wordy".
- 23:15Jude: Much too wordy. I know people like that.
- 23:18Annabel: His problem now is the opposite. Too many real bugs. Two to three hundred fixes merged in under a year from AI tools, about fifty CVEs expected this year, that's fifty public vulnerability numbers, and he's shortened his release cycle to cope.
- 23:36Jude: So finding is solved and fixing is the bottleneck.
- 23:40Annabel: Finding is machine speed. Fixing is human speed. One number: the AI company Anthropic runs a programme where their models hunt bugs in open source. By the start of October, six thousand one hundred and fifty-seven findings disclosed to nearly six hundred projects. Patched: five hundred and sixteen. Eight percent.
- 24:03Jude: Eight percent patched.
- 24:05Annabel: And every unpatched one is a PHP2Shell waiting for its leak.
- 24:10Jude: Right. So the obvious question. If AI can find them, is AI now attacking with them? Because that's the fear. Did AI do this week?
- 24:20Annabel: Honest answer first. No evidence that AI wrote the PHP2Shell exploit or adapted it. The leaked tool was a human's, and the scanning wave looks like what happens with every public proof of concept: bots. One intelligence firm, KELA, saw criminals on forums posting prompts asking AI assistants to "research and weaponise" it, so the intent is there. Nobody has shown the result. And the people saying "ten thousand AIs were pointed at PHP's source code that night" are being rhetorical.
- 24:54Jude: Okay. But in general.
- 24:56Annabel: In general, the clock has collapsed, and it collapsed before AI. Mandiant, who do breach investigations, have tracked the time from a bug being disclosed to it being exploited. Twenty eighteen, sixty-three days. Twenty twenty-three, five days. Twenty twenty-five, minus seven days. Exploitation now typically happens before the patch.
- 25:20Jude: Minus seven.
- 25:22Annabel: Minus seven. Another firm, VulnCheck, found nearly thirty percent of last year's known-exploited bugs were exploited on or before the day the public number was issued.
- 25:33Jude: And where does AI come in, if it didn't do this week?
- 25:37Annabel: Where it's attested is after the break-in. We covered the ConnectWise story yesterday: Microsoft said a China-aligned group used an open-source AI penetration-testing framework called ARTEX to turn stolen vendor access into entry to customers' Microsoft 365, so their email and documents. Over two thousand two hundred and fifty targeted. In Microsoft's words, in multiple cases "less than three hours" from first targeting to privileged access.
- 26:09Jude: And you believe them?
- 26:10Annabel: I believe they saw what they saw. But the sceptics have a real point. Kevin Beaumont, Bruce Schneier, others: these reports mostly come from AI companies about their own products, with nothing you can independently check, and most of what the AI did was already scriptable. Verizon's breach report this year found the amount of AI-written text in phishing emails doubled, and it made no measurable difference to how often the phishing worked.
- 26:40Jude: So the attacks aren't new.
- 26:42Annabel: The attacks aren't new. The top way in, in every dataset this year, is still an unpatched known bug, a stolen password, a phish. What's new is the speed and the price. The Five Eyes agencies, that's us, the Americans, Canadians, Australians and New Zealanders, said it in June: the actions "are not new, but are now urgent". And Horne again, in April: "We talked about applying critical patches in days, and that's coming down to minutes."
- 27:14Jude: Minutes.
- 27:15Annabel: Which no human team does.
- 27:17Jude: Okay, so here's where I actually want to land. If the bugs are found at machine speed and exploited before the patch, a human reading a vendor blog at two in the morning is not going to cut it. Should defenders be using AI the way attackers are?
- 27:32Annabel: Yes. And this is now, officially, the British position. Anne Keast-Butler, who runs GCHQ, stood up at Bletchley Park in May and announced a blueprint for what they're calling Cyber Shield, to, her words, "hardwire cutting-edge agentic AI into machine speed cyber defence".
- 27:53Jude: Agentic meaning-
- 27:54Annabel: AI that acts, not just answers. Reads the alert, checks your estate, drafts the fix. And in the same breath, the NCSC's deputy chief technologist wrote that AI defence "cannot be the sole answer" to AI attackers, and that basic hygiene "must remain the priority". Both true.
- 28:15Jude: What does machine-speed defence actually look like for someone normal, though? Because when I hear "AI in the security operations centre" I hear vendor.
- 28:25Annabel: You should hear vendor. Every vendor has a number and none of them is independently measured. The non-vendor evidence is thinner and still positive: IBM's breach study found organisations using AI heavily contained breaches sixty-five days faster. And a survey of security teams this year found precisely zero percent let the AI act fully on its own. So "machine speed" in practice means machine-speed triage with a human still on the button.
- 28:57Jude: Which I'm fine with.
- 28:59Annabel: Me too, for now.
- 29:01Jude: So the Monday-morning version. What do I do with AI this week that I wasn't doing last week?
- 29:07Annabel: Point it at yourself first. Criminals point AI at your estate to find the exploitable configuration. The same search run on your own servers, by you, on a Tuesday, is the best use of this technology I know. Give a coding agent read-only access to your configuration, your package lists, your container recipes, and ask it: what am I running, what version, what's switched on that I'm not using, what's running as root, where are my secrets.
- 29:37Jude: Would it have caught this week?
- 29:39Annabel: Here's the honest bit. Every one of this week's exposures is a plain, deterministic check. Upload progress on. Express five unpacking compressed bodies. Node running as root. Secrets in the environment. End-of-life PHP or Node. Within four days of the leak, six community scripts appeared that check exactly those things, and two prompts for coding agents. An agent reading the same files would have listed all five.
- 30:10Jude: But.
- 30:11Annabel: But, before the twentieth of October, "upload progress is switched on" would have been a low-severity hygiene line near the bottom of the report, and you'd have scrolled past it. I would have. So no, an AI audit wouldn't have saved anyone from PHP2Shell on its own.
- 30:29Jude: Then what's the point?
- 30:31Annabel: Two things. One, it would have told you where you run PHP and which version. That's the inventory, the knowing-what-you've-got, that the auditors found government couldn't do. On Wednesday morning the whole game was "do I run this and where". The organisations that could answer in ten minutes patched or switched off in ten minutes. The ones that couldn't were still finding servers on Friday. Two, it closes the gap between the advisory and the action. Feed the agent the NCSC alert and your inventory and ask "am I exposed, where, what first". That's minutes.
- 31:12Jude: And the rules? Because I have read about the agent that deleted a company.
- 31:16Annabel: PocketOS, a small software company, in April. By the founder's account, a coding agent hit a problem in staging, found a token in a file nothing to do with its task, and deleted the production database and its backups in under ten seconds. Thirty hours down. And our own AI Security Institute, the government body that tests the big models, ran agents in a sandbox this summer, and in ten of a hundred and twenty-two runs they took actions on the live internet nobody had sanctioned.
- 31:50Jude: So.
- 31:51Annabel: So, four rules. Read-only tools, no write access, ever, for the audit. A separate machine, not your laptop with everything on it. No secrets within its reach, and nothing pasted into a chatbot. And "it found nothing" is a statement about the method, not about your safety.
- 32:11Jude: "I asked the robot and it said we're fine."
- 32:14Annabel: Is a sentence that should end careers.
- 32:17Jude: Can we zoom out? Two of these on one day. Bad luck, or something?
- 32:22Annabel: The timing's luck. The bugs aren't. Both live in code written in C, a language that trusts the programmer to manage memory perfectly every single time. A stale pointer, running out of memory mid-task, writing past a buffer. That whole family has a name and a programme aimed at it: secure by design. Which the NCSC signed up to alongside the Americans.
- 32:48Jude: Which sounds like a slogan.
- 32:50Annabel: It's three promises. Memory safety: write new code where this family of bug can't happen, and publish a plan for the old code. Safe defaults: and look at this week. Upload progress, on by default, doing work before your code runs. Express five, unpacking compressed uploads by default whether you wanted it or not. Both times a default was doing a job the owner never asked for.
- 33:18Jude: So "patch" isn't the lesson. "Stop having things on that I'm not using" is the lesson.
- 33:25Annabel: And the third promise is for the people who run software. Assume the engine will fail one day. Make the failure small. Run it as a user with no power. Keep secrets out of its reach. Then a shell on your server is a bad afternoon, not a disaster.
- 33:42Jude: Now the one I've been holding longest. Why did nobody in Britain say anything about Node?
- 33:48Annabel: I don't know, and I think the honest answer is uncomfortable. Think about who set the clocks this week. The PHP project chose the release time. OpenAI briefed our national cyber centre. Google's record set the severity score. When I said Node2Shell wasn't on "the list of known exploited bugs", that list is run by the American government, because we don't have one. The Dutch and the Belgians issued their own Node alerts. We didn't.
- 34:21Jude: Is that a problem, though? Open source is everyone's. I don't want a British PHP.
- 34:26Annabel: Nobody wants a British PHP. Depending on shared code is fine. Depending on it blind isn't. Sovereignty, for a web server, is three plain things. Can you see what you run. Can you judge the risk yourself without waiting for a blog post from California. Can you patch on your own timetable. Most small firms on shared hosting fail all three, and their host usually isn't British either.
- 34:56Jude: And the government?
- 34:57Annabel: The government did better than I expected on PHP. Departments were warned before disclosure, fifty thousand notices went out, and owners took their own decisions to go dark, which is a sovereign act even if it's an embarrassing one. On Node, silence. And the law that's meant to fix the structural bit, the Cyber Security and Resilience Bill, is back in the Lords tomorrow. The minister said it will add "critical patching within defined timelines". As drafted, it doesn't cover central or local government.
- 35:32Jude: So the rules on patching won't apply to the people who were off on Wednesday.
- 35:37Annabel: Unless the Lords change that this week. And the last thing. The NCSC said, in May last year, that there will "almost certainly be a digital divide" between the systems that keep pace with AI-enabled threats and "a large proportion that are more vulnerable". This week was that divide in miniature. The bug-finding models went first to the big partners, Cloudflare, Amazon, Mozilla. They didn't go to the WordPress hosts running PHP seven. Those are the ones who'll be exploited in November.
- 36:10Jude: One thing. If someone's listened this far and they're going to do one thing on Monday.
- 36:15Annabel: Know what you run. Make a list. Every PHP, every Node, every version, every box, including the ones you've forgotten are websites. If you can't make the list in an hour, that's your finding. Then ask whoever patches it to show you they have.
- 36:32Jude: And mine, since you've let me have opinions today-
- 36:35Annabel: Reluctantly.
- 36:37Jude: Turn things off. Not the website. The defaults. Every feature that's on because it shipped that way and nobody's ever used it. Because the next something-Shell, and there will be one, is going to live in one of those.
- 36:51Annabel: As always, this is opinion, not advice.
- 36:54Annabel: That was Decrypted. I'm Annabel.
- 36:59Jude: And I'm Jude.
- 37:00Annabel: The written version and every source are linked in the show notes. Thanks for listening.