PHP2Shell: When AI Finds the Flaw, You Have Hours | Security Brief Issue 02
Thank you to everyone who subscribed to our first issue and to those who wrote back with ideas. Please keep them coming. If the first half of October was about attackers moving faster, the second half showed why. Artifi

Thank you to everyone who subscribed to our first issue and to those who wrote back with ideas. Please keep them coming.
If the first half of October was about attackers moving faster, the second half showed why. Artificial intelligence is now finding serious software flaws in weeks, and attackers are using AI to run intrusions at a scale and speed no human team could match. At the same time, several stories reminded us how much of our security depends on organizations we trust but do not control: cloud providers, software suppliers and the authorities that issue digital certificates.
Nagesh, Founder & CEO, Banyan Cloud
THIS FORTNIGHT IN 60 SECONDS
β’ AI found it, attackers used it. Critical βPHP2Shellβ flaws, uncovered with AI help, were probed by attackers before the fix was public and are now confirmed as exploited.
β’ An AI-run intrusion at cloud scale. Microsoft says a China-aligned group used an AI agent framework to target more than 2,250 Microsoft 365 tenants through a supplier's apps.
β’ Another cloud provider hit. Ransomware attackers breached Japan's GMO Onamae Desktop Cloud and demanded about US$50 million.
β’ Possibly the largest crypto theft yet. Binance confirmed a security incident on 29 October and paused some withdrawals. Estimates of what was taken range from US$2 billion to US$5.8 billion; none is confirmed yet.
β’ Trust broke at its root. Eight certificate authority keys from India's eMudhra could be calculated from public data, forcing emergency blocks across browsers.
β’ The quantum clock now has dates. NIST published its final guidance on moving to post-quantum cryptography.
THREAT RADAR
The four developments we think matter most, and what to do about each.
1. PHP2Shell: patch windows are now measured in hours
What happened. On 21 October, the PHP project released emergency fixes for two critical flaws (CVE-2026-109618 and CVE-2026-109619) that allow attackers to run code on a web server without logging in. Researchers at Calif, Wiz and OpenAI each reported the problem independently, and Wiz says AI models helped it find and test the bugs. Memory-safety flaws of this kind have been sitting in PHP for about a decade. Amazon saw an exploitation attempt in its honeypots on 17 October, four days before the fix was public, and widespread attempts began within hours of a proof-of-concept leaking. The EU's Commission, ENISA, CERT-EU and Europol issued a joint call to patch, and the UK's NCSC advised taking unpatched systems offline. On 24 October a researcher reported a successful attack on a test server, and on 26 October CISA added both flaws to its Known Exploited Vulnerabilities catalogue, giving US federal agencies until 29 October to patch.
Why it matters. PHP runs a large share of the web, including WordPress, Drupal and Joomla, and it is often bundled inside appliances and third-party products where teams don't know it exists. One estimate puts about 79% of internet-facing PHP services on versions that no longer receive updates. AI is now finding old, deep flaws quickly, and the time between disclosure and attack has nearly disappeared.
What to do. Find every PHP runtime you have, including inside containers, appliances and vendor products. Update to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35, starting with internet-facing servers. If you can't patch within hours, restrict access and confirm your web application firewall is blocking, not just monitoring. Check for compromise from 17 October onwards and rotate secrets held by PHP applications. Don't run unvetted βproof-of-conceptβ code from GitHub, as some contained malware.
2. Storm-0940: an AI agent, a supplier's apps and 2,250 cloud tenants
What happened. On 21 October, Microsoft attributed an attack on ConnectWise's Cloud Backup service to Storm-0940, a China-aligned group. Using credentials stolen in that breach, the attackers signed in to Microsoft 365 through ConnectWise applications that customers had already authorized, so no user ever had to log in. They used ARTEX, an open-source AI agent framework, to automate everything from reconnaissance to data theft. Microsoft saw sign-in attempts against more than 2,250 tenants, at up to 2,900 attempts a minute. Most failed, but some attackers gained privileged cloud access in under three hours.
Why it matters. This is what an AI-assisted supply-chain attack looks like in practice. Many organizations protect their people with MFA, but the apps they connect to their cloud often hold standing access with far less scrutiny. When a supplier is breached, every customer who granted that supplier access is exposed.
What to do. If you use ConnectWise cloud applications, follow Microsoft's guidance: keep the affected connections disabled until ConnectWise says otherwise, search your tenant using the affected application IDs rather than display names, and preserve your logs from at least 1 October. More broadly, see this issue's Cloud Risk Spotlight.
3. Cloud providers: a second attack, and lessons from the first
What happened. On 20 October, GMO Internet detected an intrusion into Onamae Desktop Cloud, its virtual server service in Japan, and said malware may have been installed on up to 26,739 customer servers. GMO later confirmed data was stolen and that attackers demanded Β₯8 billion (about US$50 million), which it says it will not pay. On 27 October it asked customers to disconnect and shut down their computers as a precaution. Since then, data that appears to relate to customers has been published, GMO has offered a reward of up to US$1 million for information identifying the attackers, and it says it will not restart the service until its security has been rebuilt. It has found no sign of ransomware on the customer devices it has analyzed so far. Separately, Sony Network Communications disclosed on 27 October that attackers had accessed the member database of its So-net and Plala internet services, exposing records of about 2.58 million current and former customers. Meanwhile, IDC Frontier gave an update on the IDCF Cloud attack we covered in Issue 01: backups could be recovered for 194 of its 495 affected customers, all of whom had used its snapshot feature.
Why it matters. Two Japanese cloud providers were hit by ransomware in a single month, and a major Japanese internet provider lost millions of customer records. In the IDCF case, whether a customer could recover depended largely on a backup choice made long before the attack. Customers with no usable copy had little they could do.
What to do. Whatever precautions your cloud provider takes, protecting your data remains your responsibility as a customer. Keep at least one backup copy outside your provider's platform, ideally one that cannot be altered or deleted, and test restoring it regularly. Know which of your services would stop if one provider went offline for a week, and have a plan to run them elsewhere. Keep a copy of your cloud configurations so you can rebuild, and protect your own console access with strong MFA.
4. SafePay: 600 victims, mostly through familiar doors
What happened. On 20 October, the FBI, CISA and partner agencies in Germany and the UK published a joint advisory (AA26-293A) on SafePay, which they call one of the most active ransomware groups of 2026. It has breached at least 600 organizations in the US and Europe since August 2024, including critical infrastructure. Its way in is usually password guessing, phishing or known flaws in Cisco, SonicWall and Veeam products, some of them several years old.
Why it matters. As with FortiBleed in our last issue, the entry points are not new. Remote access devices and backup servers with old vulnerabilities or weak passwords remain the most reliable way into an organization.
What to do. Check that the vulnerabilities named in the advisory are patched, enforce MFA on all remote access, and watch for remote-support tools such as AnyDesk or TeamViewer appearing where they shouldn't.
REGULATOR WATCH

REGIONAL SPOTLIGHT: MIDDLE EAST
High attack volumes, and a growing market to meet them
UAE: On 20 October, the UAE Cyber Security Council said it had detected and stopped nearly 750,000 cyberattack attempts against vital sectors, particularly energy and education. They included remote code execution, DDoS, phishing and AI-supported attacks. The Council stressed that a high number of attempts does not mean breaches succeeded, and pointed to its national system of continuous monitoring, early detection and coordinated response.
Saudi Arabia: A National Cybersecurity Authority report published on 14 October puts the Kingdom's cyber security market at SAR 16.3 billion in 2025, growing about 10.7% a year since 2023. The private sector accounts for 69% of spending.
REGIONAL SPOTLIGHT: INDIA
Banking customers in the crosshairs, and a test of digital trust
State Bank of India: On 21 October, SBI disclosed that 3,897 customer records had been found among about 375,000 records circulating on the dark web. It was alerted on 16 October by CERT-In, the Defense Cyber Agency and its threat intelligence provider. The exposed data includes names, mobile numbers, dates of birth, addresses and customer IDs, but not passwords, PINs or OTPs. SBI says customer funds and operations are unaffected and has engaged a CERT-In-empaneled agency to investigate. The source of the leak has not been disclosed.
CyberDakait: Research from Group-IB, reported on 29 October, describes an Android malware operation targeting Indian banking customers since July 2025. Callers posing as bank staff persuade victims to install fake apps that steal banking credentials and card details and take remote control of the phone. About 17,960 devices have been compromised and more than 7.6 million SMS messages intercepted, with 11 Indian banks and two government programmes impersonated.
eMudhra: After eight of its certificate authority keys were found to be calculable from public data (see Quantum-Ready Corner), MeitY formed a team to investigate and CERT-In issued advisory CIAD-2026-0050. Officials say eMudhra's Indian eSign and digital signature services, which are regulated separately, are unaffected. On 22 October eMudhra announced it will exit the global SSL/TLS certificate business.
CLOUD RISK SPOTLIGHT
The apps that never log out
Every time someone connects a backup tool, a CRM integration or a productivity add-on to Microsoft 365 or Google Workspace, they create an identity for that app. Many of these hold broad, standing permissions, such as reading all mail or all files. They sign in without MFA prompts, and they are rarely reviewed once approved. The Storm-0940 attack used exactly this route.
Five checks for this fortnight
1.List every third-party app with access to your cloud tenant, and who approved it.
2.Remove apps nobody uses and cut permissions that go beyond what each app needs.
3.Stop ordinary users from granting broad app permissions without an administrator's approval.
4.Apply conditional access and location limits to app identities, not just to people.
5.Monitor sign-ins by apps (not only by users), and keep those logs long enough to investigate.
QUANTUM-READY CORNER
Two lessons in cryptographic trust
On 27 October, NIST released its final report on moving to post-quantum cryptography (NIST IR 8547). It sets out which of today's algorithms are vulnerable and when they may no longer be allowed in US federal systems. Under the June 2026 executive order, federal high-value systems must move their key exchange to quantum-safe methods by 2030 and their digital signatures by 2031, with 2035 the broader target. NIST advises organizations to prioritize data that must stay secret for many years, to consider their whole cryptographic supply chain, and to look at hybrid approaches that combine today's encryption with post-quantum methods.
The same fortnight brought a reminder that cryptography can fail without any quantum computer. Eight intermediate certificate authority keys belonging to India's eMudhra had been generated with too little randomness during a 2025 key ceremony, so anyone could calculate them from public certificates. One researcher did it in under a second. Browsers and operating systems blocked them within days and removed emSign roots from several trust stores. There is no evidence the keys were misused. (India's response is covered in this issue's Regional Spotlight.)
Both stories point to the same first step: knowing what cryptography you use and whom you trust. That inventory, sometimes called a cryptographic bill of materials (CBOM), is what lets you move quickly when an algorithm or a certificate authority can no longer be trusted.
Three questions to ask your team
- Which certificate authorities do our websites, applications and containers trust, and could we replace one within a week?
- Do any of our applications or devices carry their own fixed list of trusted certificates that won't update automatically?
- Have we mapped which systems will need post-quantum key exchange first, against NIST's timelines? HYGIENE TIP OF THE FORTNIGHT
Think before you click βAcceptβ
When an app asks to βread your mailβ, βaccess your filesβ or βact on your behalfβ, it may keep that access long after you've stopped using it. Only approve apps you recognize and need. Every few months, open the connected apps page in your Microsoft, Google or social media account and remove anything you no longer use.
THE BANYAN CLOUD VIEW
Your security perimeter now includes everyone you trust
Look at this fortnight's stories and notice how few of them started inside the victim's own walls. Microsoft 365 tenants were reached through a backup supplier's apps. Hosting customers were hit through their cloud provider. Websites around the world briefly depended on certificate keys that anyone could calculate. In each case, organizations were exposed by something they had trusted, often years ago, and had not looked at since.
In our first issue we argued that security has to start with data: knowing what you hold, where it lives and who can reach it. This fortnight adds a fourth question: whom have you trusted with it? Every connected app, cloud provider, supplier and certificate authority is part of your data's security, whether it appears on your risk register.
AI is making all this faster. It found PHP2Shell, and it drove the Storm-0940 intrusion. Defenders can use it too, but only if they have a clear, current picture of their data, identities and dependencies to work from. Without that picture, faster tools just reach the wrong conclusions sooner.
FROM BANYAN CLOUD
Recap: our webinar with the Indo-American Chamber of Commerce
Nagesh, our Founder & CEO, recently led a webinar for the Indo-American Chamber of Commerce (IACC) on the security challenges business leaders face today. He covered:
β’ AI risks: phishing, deepfakes and shadow AI
β’ Cloud security: excessive access, misconfigurations and third-party risk
β’ Cross-border compliance: incident reporting, data protection and AI governance
β’ Business resilience: building trust through continuous security and compliance evidence
Many of these themes run through this issue, from third-party apps in the cloud to AI-driven attacks. Thank you to IACC and to everyone who joined.
Webinar date: Oct 28th, 2026
Link to recorded session: https://www.youtube.com/watch?v=PLrgcTHtBhk
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.