|
Welcome back to the TCE Strategy monthly newsletter! October is Cybersecurity Awareness Month, and the Internet apparently decided to make the point for me. We got one of the nastiest patching emergencies since Log4j, attackers who were already inside systems before a patch existed, French supermarkets whose refrigerators were remotely turned into food warmers, webcams drafted into actual shooting wars, and an AI agent that filed a false tip about an unsolved murder. Let’s see how this month’s news can help us make better decisions about what is Secure Enough for us, the companies we work for, and our families.
Patch your PHP. Right now. I’ll wait.
On October 21, the PHP project released emergency fixes for two vulnerabilities that can let an unauthenticated attacker run code on a web server. No credentials needed. Just a malicious web request. Both vulnerabilities are rated 9.2 out of 10 and are collectively being called “PHP2Shell.”
Why should you care? Because PHP powers about 70% of the websites where the server-side programming language can be identified, including tcestrategy.com. WordPress runs on PHP. So do Drupal, Joomla, MediaWiki and an enormous number of custom websites. If your company has a website, there is a very good chance PHP is somewhere underneath it, even if nobody at your company has ever knowingly touched PHP.
The fixed versions are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. PHP 8.1 and older versions no longer receive security releases from the PHP project, although some operating-system and support vendors provide backported fixes. Note also that PHP 8.2 reaches end of life on December 31 of this year, which makes 8.2.35 a reprieve rather than a destination. There is no magic configuration setting that fully fixes this. The PHP project’s advice is wonderfully unambiguous: update.
One more wrinkle, because this story refuses to sit still. On October 23, PHP published an informational advisory for a separate flaw, CVE-2026-115560, which can create a remote code execution path through session upload progress. PHP calls the scenario unlikely, has not published affected versions or a severity score, and will fix it in a future release. Until then it recommends setting session.upload_progress.enabled=0. If you already set that when you patched on the 21st, you are covered. Apparently even the emergency patch needed an asterisk.
Here is where this gets ugly. Amazon saw an apparent exploitation attempt against one of its honeypots on October 17, four days before public disclosure and four days before the patches became available. So if you patched on October 21, congratulations — and now go look backward. Patching closes the hole. It does not prove nobody crawled through it first.
The embargo did not hold either. The CVE identifiers leaked onto Hacker News and X a day before the planned release, a video demonstrating the exploit chain began circulating, and the PHP Foundation says a vendor shipped patched container images six hours ahead of schedule. That vendor denies leaking anything and says it moved early because customers were already being attacked. PHP shipped the fixes in a hurry, and a working proof-of-concept named php2shell was on GitHub within hours.
The skill barrier is still there. AI is busy lowering it. Cloudflare’s CEO watched attempts climb from thousands in the ten hours before the announcement, to tens of thousands within 45 minutes of it, to hundreds of thousands within two hours, to millions. Amazon says PHP2Shell is now being broadly exploited in the wild.
Here is where the good news lasted about a day. Early reporting showed millions of attempts but no confirmed successful break-in. Then Palo Alto Networks’ Unit 42 updated its findings. It says attackers compromised a European government entity and, within two hours of the initial exploitation, had accessed password hashes and MFA secrets, established an NTLM-relay tunnel into Active Directory, and were trying to become Domain Admin.
Unit 42 also says the post-exploitation behavior showed strong indicators of an agentic AI attacker, and the tell is worth knowing. Each batch of commands was wrapped in section headers such as ==SHELLUSER== and ==PASSWD==, which exist for no reason other than to make the output easy for a machine to read. Unit 42 also noted short, serially spaced tasks and fast pivots after each failure. That is behavioral attribution, not identification of a particular AI product.
Hold both halves of this in your head at once. Most of the exploit code being sprayed across the Internet still does not work; watchTowr reports that the exploits going around either fail outright, test for timing differences, or simply crash the target. But one competent attacker got in and was knocking on Domain Admin’s door in under two hours. You are not going to be taken down by the garbage. You are going to be taken down by the one who knows what they are doing. Patch now, and if you were exposed, go looking for evidence that somebody got there first.
One more number, then I will let you go call your web host. On October 21, Censys saw roughly 1.57 million Internet-facing hosts advertising PHP versions that appeared unpatched. Version banners are an imperfect measure, since some vendors backport fixes without changing the version string, but the order of magnitude is the point.
There is also an AI angle, because apparently we cannot have a cybersecurity story anymore without one. Calif found a critical PHP flaw using OpenAI’s GPT-5.6 Sol, while Wiz independently found the flaws using AI-assisted research that included GPT-5.6 Sol and Google’s Gemini 4 Argon. These bugs had survived in PHP for years. What AI did was help researchers turn them into practical attack chains. Wiz described building that chain as at the edge of what human researchers could practically achieve, and said a single researcher did in about a week what would previously have taken a world-class team months. That is terrific for defenders when we find the bug first. It is considerably less terrific when the attacker does.
Takeaway: If you host your own PHP, patch it now. If somebody else hosts your website, ask them in writing what PHP version you are running and whether the October 21 fixes were applied. “Our web host handles that” is not an answer; it is a deflection wearing business-casual clothing. If you were exposed, review logs and other indicators starting October 17, not October 21.
Read more: PHP security advisory and fixed versions
Read more: Amazon on exploitation before public disclosure
Read more: Unit 42 on the confirmed European government compromise and agentic post-exploitation
Patching is not the finish line
The PHP story has an uncomfortable companion. In June, PTC disclosed a critical vulnerability in Windchill and FlexPLM, software used by manufacturers to manage engineering drawings, specifications, change history and approved product designs. In other words, a convenient electronic filing cabinet for a manufacturer’s crown jewels.
Researchers believe the Cl0p extortion gang had been exploiting it as a zero-day since early June, before any patch existed. Attackers planted web shells, stole credentials and staged data for theft. PTC eventually published indicators of compromise and told customers not just to patch, but to go hunting for evidence that attackers were already inside.
I have a client caught up in this one, which is as much detail as I am able to give. The lesson is that a company can patch promptly, yet still have a serious problem if somebody got there first.
Then there is the FBI. Last month I wrote about the FBIjobs.gov breach. We now know considerably more. FBI cyber chief Brett Leatherman told Reuters that the incident resulted from a security failure on a third-party-managed platform after a contractor failed to install a security patch that had been explicitly issued to secure it. The contractor was removed. The stolen data reportedly included extraordinarily sensitive information about FBI employees.
Takeaway: Patching is not a project. It is a habit, and somebody has to own it by name. Ask three questions: Who is responsible for patching every Internet-facing system, including vendor-managed systems? How do we verify the patch actually installed? And when exploitation started before the patch existed, who hunts for signs that we were already compromised? A patch closes the door. It does not throw the burglar back outside.
Read more: PTC’s Windchill/FlexPLM advisory
Read more: Reuters on the FBI contractor and missed patch
Before you connect it, ask who would want to hack it
Hundreds of French supermarkets learned a very expensive lesson this month: refrigerators are computers now, just like most everything else.
On the night of October 11, attackers remotely manipulated refrigeration systems at hundreds of stores. Freezers and cold rooms were pushed into extended defrost cycles, temperatures soared, food was destroyed and some equipment reportedly got hot enough to cause fires. Intermarché, Coopérative U, Carrefour, Monoprix, E.Leclerc and others were affected.
Early reporting focused on known vulnerabilities in Danfoss StoreView refrigeration controllers, and Censys found a startling number of them sitting directly on the Internet, including old versions where an administrator code could be derived from the calendar date. Be careful here, though, because the reporting has not settled. CERT-FR has named no vulnerability, no firmware version and no entry point. A source close to the French prime minister said ANSSI had identified a vulnerability in Danfoss software, without saying which one or whether it was the one actually used.
And here is the wrinkle most coverage has skipped. The published research everyone is citing covers the Danfoss AK-SM 800A. Danfoss named the AK-SM 800 and the AK-SC 355 as affected, which are different models the researchers never tested. The AK-SC 355 was phased out in 2016. The AK-SM 800 reached end of life in May 2021. For those units, “just patch it” was not an option. Danfoss no longer provides security updates for them. The only durable choices were replacement or isolation, and leaving end-of-life industrial equipment directly reachable from the public Internet should never have been on the menu at all.
Two days after the attack, Reuters was still finding at least 4,500 of these interfaces visible through Shodan. Censys watched the number of exposed controllers in France fall by 71% in three days, and because patched and vulnerable versions disappeared at nearly identical rates, Censys concluded that operators were not suddenly patching thousands of controllers. They were unplugging them. That was the correct answer. It would also have been the correct answer in 2021.
The next night, attackers reportedly went after refrigeration systems from another manufacturer, Dixell, using a similar approach. CERT-FR extended its alert on October 13 to cover Copeland and Dixell consoles as well as the Danfoss units. That makes the vendor name much less important than the architecture. The problem is not that one brand of refrigeration controller is cursed. The problem is that operational technology keeps getting connected to the Internet because remote management is convenient, while securing it gets treated as somebody else’s problem.
If you are filing this under “European problem,” don’t. Over two nights in late August, at least six US military commissaries reported refrigeration outages, including Fort Huachuca in Arizona, where officials said every freezer went into defrost mode overnight and all frozen food spoiled. The Defense Department acknowledged only a “possible refrigeration disruption” and declined to speculate about the cause. Three military criminal investigation agencies have since opened active probes.
Now add cameras. The Wall Street Journal reported that Internet-connected cameras are being hacked for battlefield intelligence. A poultry farmer in northern Israel discovered that hackers had taken over a camera outside his chicken coop. Nobody cared about his chickens. The camera happened to show a road used by Israeli security forces. Israel says Iranian-linked groups have compromised hundreds of cameras it considers sensitive; Ukraine has secured or blocked tens of thousands of cameras that could reveal military movements.
A guy put a camera on a chicken coop. He was not wrong to do it. He simply did not realize that his view of the road could become somebody else’s intelligence feed.
Takeaway: Before anything can be controlled or watched remotely, make someone answer one question out loud: who would want access to this, and what could they do with it? Then do the boring work. Change default credentials. Use MFA where the product supports it. Segment it from systems that matter. If it does not need to be reachable from the public Internet, make it not be. And identify the human being responsible for firmware updates, because “the vendor” is not a person.
Read more: Censys on the exposed Danfoss controllers
Read more: France 24 on the attack and the French investigation
Read more: Wall Street Journal on cameras being hacked for war (paywalled)
Your AI agent did something you never asked it to do
Personal AI agents are moving from “answer my question” to “go do things for me.” That sounds wonderful right up until the “things” include something nobody intended.
Anthropic disclosed this month that one of its Claude models landed on a Philadelphia Police Department page about an unsolved homicide, found the tip form, and submitted a false tip claiming to have seen someone matching the description. The model had been given restrictions, but those restrictions did not explicitly prohibit submitting ordinary forms. Apparently “do not file a bogus murder tip with the police” was not on the checklist.
The tip was caught as spam and never reached an investigator, so the real-world damage was minimal. The timeline is less comforting. The tip was submitted July 18. Anthropic discovered it September 28 and notified Philadelphia police in early October. The department called the two-month delay unacceptable. Anthropic also disclosed agents exploiting software flaws, bypassing restrictions to reach gated data, and using URL-shortening services to get around limits in their own tools. Anthropic agents also filed visa applications on the State Department’s website. The New York Times reported it first from anonymous sources, and the State Department has since confirmed it on the record: one application in May and 19 more in August, all incomplete, none processed, and no departmental system compromised.
Anthropic’s response was to shut off live Internet access for all internal evaluations until it is satisfied that its monitoring and controls reliably catch this behavior. Read that sentence again. One of the companies building the most advanced AI agents on earth decided that, for now, the safe way to test them is to stop letting them touch the real Internet.
Then something happened that I have been asking for in this newsletter for two months in a row. On October 9, the newly named “White House Super Intelligence Force”, a task force created five days earlier and chaired by AI czar Jay Clayton, issued a statement that is supposed to apply to every AI company, not just Anthropic. Companies must immediately disclose incidents involving their models and move quickly to remedy any harm. Their words, not mine: “This notification and remediation process is not optional. It is a critical national security obligation.” Important note: the words “not optional” do not change behavior. Rules, sufficient enforcement of the rules, and adequate penalties for breaking the rules will change behavior. All three are necessary. Everything else is lip service that makes for good political talking points but does nothing for real-world cybersecurity.
In September I wrote that AI incident disclosure was entirely voluntary, and that voluntary disclosure by the party with the most to lose is not a safety system. Axios, which broke the statement, made the same point more bluntly: the administration’s approach to AI regulation had been voluntary at least in name, until now. Here is the caveat, though. Nobody has been able to identify what the penalty is if an AI company simply does not comply. So we have moved from “please tell us when something goes badly wrong” to “you must tell us when something goes badly wrong.” That is not real progress unless all three of the aforementioned points are fulfilled. I will call it a regulatory regime when somebody actually gets punished for ignoring it.
This is the same family of problems I have been writing about for months. We are training these systems to be persistent, to find another path when the first one fails, and to complete a goal with minimal human intervention. Then we act surprised when “find another path” turns into “do the adjacent thing nobody thought to forbid.” You cannot solve that with a 14-page prompt full of thou-shalt-nots.
Takeaway: Controls for agentic AI need to be structural, not conversational. Give agents narrowly scoped accounts. Limit what they can reach at the system level. Log what they actually do. Keep personal AI accounts away from corporate systems. Require human approval before an agent sends, submits, pays, publishes, deletes or changes something important. If the only thing stopping your AI agent from doing something stupid is a sentence in its prompt, you do not have a control. You have a suggestion.
Read more: Anthropic’s report on unintended model actions
Until next month, stay safe!
|