I Took a Week Off and the Internet Got Hacked
I took a week off from AI, work and YouTube. A little mini vacation. I came back to what felt like the whole internet being hacked: compromises, breaches, finger pointing, some of it AI related and some of it not. I spent a few hours catching up, got overwhelmed, and recorded a video walking through the stories that I think matter most for where AI is going and how people will see it.
None of this is my own testing. These are the posts and articles I read, with links to all of them below, plus my take.
AI agents inside bank accounts
The thing that pulled me into the research loop was a post from Madison Mills on an Axios report: major US banks are detecting unauthorized third-party AI agents inside customer accounts, including some that appear to act independently of the user's instructions. Bank executives told Axios they struggle to tell human users from agents, and to tell whether an agent was authorized by the customer at all. Some are planning for worst-case and "day after" scenarios.
I've been around financial institutions before, and from that perspective I'm not surprised. A lot of banking software runs on old systems, outdated operating systems and proprietary code with slow patching and testing. They were already fighting a hard battle. Agents moving faster than they can respond makes it close to impossible. That alone would be headline news. This week it got buried.
Japan, and French supermarket fridges
On October 22, cyberattacks hit Japanese police and businesses, with major companies forced to suspend online services. A lot of the online buzz claimed internet access across Japan was being shut off.
Then one I missed from before my break. On October 11, attackers got into the refrigeration controllers of roughly 700 French supermarkets and set them to defrost overnight. Stores had to throw out spoiled stock, with damage estimated at tens of millions of euros.
How does this connect to AI? It keeps getting easier for anyone, skilled or not, to chain together AI tools to find and hit weak targets. Internet-connected microcontrollers are convenient, but they sit in awkward places, rarely get updated and sometimes can't be. That was already a problem for power plants and processing plants. AI makes those targets easier to reach.
The NSA told everyone to patch PHP or take it offline
The post going viral when I recorded was from NSA Cyber: effective immediately, organizations using PHP should prepare for an update addressing two critical vulnerabilities, and if they can't update right away, take the system offline. They followed up with specific mitigation guidance and an official NSA advisory. The two CVEs are CVE-2026-109618 and CVE-2026-109619.
People have joked that PHP has been dead since 1995, but a huge amount of the web still runs on it, including WordPress. When the US government steps in and tells you to turn something off, it's serious.
The part I keep thinking about is disclosure. The bugs, now called PHP2Shell, were found back in August. Security firm Calif found the first one with help from OpenAI's GPT-5.6 Sol model and reported it to PHP on August 11, and Wiz Research found and reported them independently on August 24. Then came the normal process: an embargo while the PHP team built a fix and coordinated the release, with public disclosure planned for Thursday, October 22.
On October 20, two days before that date, a post on Hacker News claimed two critical PHP flaws were about to drop. It was taken down, but screenshots spread, and researchers at watchTowr confirmed the leak was real and reproduced one of the bugs. PHP pulled its release forward to October 21 at 14:00 UTC. According to Sansec's timeline, Cloudflare saw early attacks at 07:43 UTC that morning, before the patch. Amazon says its honeypots saw an exploitation attempt on October 17, even before the leak, and how that attacker got the details is still unknown.
So the gap between "the bug is out there" and "there's a patch" wasn't a few hours. Counting from the leak, it was the better part of a day, and some attackers had it earlier. That's exactly what coordinated disclosure is supposed to prevent, and this time it isn't just skilled attackers who can act on a leak. Someone with very little skill and a capable AI can too.
There's a patch now. But a lot of PHP runs underneath old shared systems, and someone has to know who owns each server and get it updated. Working in tech, I've seen how much effort that takes, and I'm glad people are taking it seriously.
A certificate authority's private key was the number 4
This one is almost comical. eMudhra is a certificate authority, one of the organizations the internet trusts to sign the certificates behind HTTPS. Scott Helme's write-up explains it best: one of eMudhra's CA private keys was literally the number 4. Not a long number that starts with 4. A key that should be a random 384-bit number was 47 bytes of zeros and then a single byte, and all eight affected keys were like that.
Browsers moved fast. Updated Chrome, Firefox and Safari are fine, so update your browser. Tools like curl, Python and Node.js that rely on a CA bundle on disk are a different story until those bundles are updated. eMudhra has since announced it is exiting the global SSL/TLS certificate business, which Reuters reported as well.
This isn't an AI story exactly. It's AI adjacent. But it shows how much of the internet rests on trust that nobody double checked.
GitLab got a CVSS 10.0
Then GitLab, the self-hostable alternative to GitHub, got a critical flaw scored CVSS 10.0, the highest score possible, affecting about 12 years of releases. GitLab shipped a critical patch release. If you self-host GitLab, update.
This post from Andras Bacsai summed up the week better than I can: six security headlines in one list, plus a SpongeBob "ight imma head out." That's honestly how I feel.
What I think comes next
I'm torn. These vulnerabilities always existed. AI is exposing them and making them easier to exploit, and ideally that leads to a safer internet. But the next 12 to 24 months are going to be nonstop: hacks, leaks, exploited systems, and privacy getting challenged along the way.
Meanwhile the strongest models are being handcuffed for defenders too. Unless you're in the club with the frontier labs, a small company is not getting access to the best security models. So it becomes a cat-and-mouse game, amplified by open-weights models, adversaries in other countries and how old a lot of this infrastructure is.
It's getting real. My worry is that AI gets treated as a global threat and open weights get banned. The optimistic version is that this blows over, things get patched, and we end up somewhere better. I'm going back to local AI on my new Mac Studio, and to making my own setup a little more secure.
What did I miss? Tell me the hacks and CVEs I didn't cover.
Sources
- Madison Mills on the Axios report about AI agents in bank accounts
- Cyberattacks Hit Japanese Police and Businesses, News On Japan
- 2026 French supermarket refrigerator hack, Wikipedia
- NSA Cyber's PHP warning and mitigation follow-up
- NSA Advisory: Critical Patches Recommended for PHP Vulnerabilities
- CVE-2026-109618 and CVE-2026-109619 on NVD
- Eight Bits of Trust: eMudhra's CA keys cracked in milliseconds, Scott Helme
- eMudhra to exit the global SSL/TLS certificate business
- Indian firm eMudhra exits global digital certificates business, Reuters
- Critical GitLab flaw, The Hacker News
- GitLab Critical Patch Release 19.5.1
- The internet this week, Andras Bacsai
- PHP2Shell timeline, Sansec
- PHP2Shell write-up, Wiz Research
- PHP2Shell: what we know about the leak, Check Point
- Amazon Threat Intelligence on PHP2Shell exploitation before disclosure
🎥 Gear links
- Apple Mac Studio M5 Max
- Sony a6700 with 16-50mm kit lens
- Shure SM7B dynamic microphone
- Shure MVX2U Gen 2 audio interface
- Audio-Technica ATH-M50x headphones
- ADAM Audio D3V speakers
- GMKtec EVO-X2 AI Mini PC
- Logitech G PRO X Superlight 2 mouse
- NEEWER NL288 LED video light
Affiliate links may earn me a commission at no extra cost to you. As an Amazon Associate I earn from qualifying purchases.
