An AI Hacker Used Claude and DeepSeek to Breach 600 Firewalls. That Wasn't Even the Worst Thing This Week.
- Patrick Duggan
- Feb 23
- 5 min read
title: "An AI Hacker Used Claude and DeepSeek to Breach 600 Firewalls. That Wasn't Even the Worst Thing This Week."
date: 2026-02-23
author: Patrick Duggan
tags: [cybersecurity, threat-intel, fortigate, beyondtrust, ransomware, stix, cisa-kev]
# An AI Hacker Used Claude and DeepSeek to Breach 600 Firewalls. That Wasn't Even the Worst Thing This Week.
**A Russian-speaking threat actor used commercial AI to pop 600+ FortiGate firewalls in 55 countries. A CVSS 9.9 pre-auth RCE is being used in ransomware campaigns. Mississippi's only Level I trauma center shut 35 clinics. PayPal leaked SSNs for six months. Cloudflare deleted its own IP prefixes. And CISA added 13 new KEVs in February. This is one week.**
The FortiGate AI Campaign
Amazon Web Services dropped the report on February 18. A financially motivated, Russian-speaking threat actor breached 600+ FortiGate appliances across 55 countries in five weeks — and they did it with AI.
The tools: **DeepSeek** and **Anthropic Claude** (yes, the same Claude you might be reading this blog post through). Plus custom orchestration: ARXON, a Python MCP server, and CHECKER2, a Go-based automation tool.
The method wasn't a zero-day. It was embarrassingly simple: the attacker scanned management interfaces on ports 443, 8443, 10443, and 4443, then authenticated with **weak and reused credentials**. Single-factor auth. In 2026.
Post-access, they used CVE-2019-6693 — a **seven-year-old** Fortinet password decryption vulnerability — to extract full device configurations. Then Active Directory compromise. Then Nuclei vulnerability scanning of internal networks. Then attempts to access backup infrastructure — the classic ransomware preparation playbook.
One deployment in the campaign targeted **2,516 systems across 106 countries**.
The AI didn't find the zero-day. The AI didn't break the encryption. The AI helped the attacker **move faster** through post-exploitation — analyzing configs, identifying targets, automating the tedious work that used to require a team. The attacker was a single individual.
This is the future of cybersecurity: not AI finding new vulnerabilities, but AI making **existing human laziness** exponentially more exploitable.
**Are your FortiGate management interfaces exposed to the internet with single-factor auth?** Check. Now.
BeyondTrust CVE-2026-1731: CVSS 9.9
On February 6, BeyondTrust disclosed CVE-2026-1731 — a pre-authentication remote code execution vulnerability in Remote Support (RS) and Privileged Remote Access (PRA). CVSS 9.9. An unauthenticated attacker can execute arbitrary OS commands via specially crafted requests. No credentials needed. No interaction required.
CISA added it to the Known Exploited Vulnerabilities catalog on February 13 and activated the **ransomware campaign flag**. This is being used in the wild right now.
Palo Alto's Unit 42 confirmed exploitation: web shell deployment, C2 using VShell and SparkRAT, lateral movement, data exfiltration. Targeted sectors: **financial services, legal, healthcare, higher education, technology**. Targeted countries: **US, France, Germany, Australia, Canada**.
BeyondTrust's SaaS instances were auto-patched on February 2 (before disclosure). Self-hosted customers need to upgrade RS to 25.3.2 or PRA to 24.3.5.
**Federal remediation deadline: March 6, 2026.**
If you run BeyondTrust RS or PRA self-hosted, and you haven't patched, you are being actively targeted by ransomware operators. That's not a prediction. That's CISA saying it.
Mississippi's Only Trauma Center: 35 Clinics Dark
On February 19, ransomware hit the University of Mississippi Medical Center — the state's only academic medical center and only Level I trauma center.
**All 35 UMMC clinics statewide closed.** Elective surgeries canceled. The EPIC electronic medical record system went down. Staff reverted to pen and paper.
Emergency departments stayed open — Jackson, Grenada, Madison County, Holmes County. The kidney dialysis clinic at Jackson Medical Mall kept running. Everything else: dark.
The ransomware group hasn't been publicly identified. UMMC confirmed contact with the threat actor. FBI and DHS are engaged. No word on ransom payment.
This is Mississippi's healthcare backbone. The state's poorest patients. The state's most vulnerable populations. 35 clinics, closed, because someone found a way in.
PayPal: Six Months of SSNs
PayPal disclosed on February 10 that a **coding error** in the Working Capital loan application exposed customer data from **July 1 through December 13, 2025**. Five and a half months.
Approximately 100 PayPal Working Capital business customers had names, email addresses, phone numbers, business addresses, **Social Security numbers**, and dates of birth exposed. Some customers experienced unauthorized transactions. PayPal issued refunds.
This wasn't a hack. This wasn't a zero-day. This was a **code error** that sat in production for 166 days. No CVE was assigned because it wasn't a vulnerability in the traditional sense — it was a bug that nobody caught.
PayPal is offering two years of three-bureau credit monitoring through Equifax. If your SSN was exposed for six months, you'll need more than that.
Cloudflare: Self-Inflicted
On February 20, Cloudflare experienced a 6-hour, 7-minute global outage. Not an attack. A **BYOIP prefix deletion bug**.
A cleanup task meant to remove prefixes flagged as `pending_delete` had a bug: when the parameter was passed without a value, the API interpreted it as an empty string and **queued ALL 1,100 BYOIP prefixes for deletion** instead of only the flagged ones. Out of 4,306 total BYOIP prefixes, 25% were withdrawn.
Services affected: CDN, Spectrum, Dedicated Egress, Magic Transit, the 1.1.1.1 DNS resolver.
800 prefixes were restored automatically within a few hours. The remaining 300 required manual restoration because the service configurations had been deleted from the edge. Full recovery at 23:03 UTC.
The buggy code was merged on February 5 but didn't execute until February 20. Fifteen days of a time bomb sitting in production. Part of Cloudflare's "Code Orange: Fail Small" resilience initiative.
Irony is dead.
Roundcube: Weaponized in 48 Hours
CISA added two Roundcube Webmail vulnerabilities to the KEV catalog on February 20:
- **CVE-2025-49113**: Deserialization of untrusted data. CVSS 9.9. Remote code execution by authenticated users.
- **CVE-2025-68461**: XSS via SVG `animate` tag. CVSS 7.2.
Attackers weaponized the deserialization vulnerability **within 48 hours of public disclosure**. Roundcube is widely deployed in government, enterprise, and hosting environments.
**Federal remediation deadline: March 13, 2026.**
February's CISA KEV Scorecard
| Date | Count | Highlights |
|------|-------|------------|
| Feb 3 | 4 | Sangoma FreePBX, GitLab SSRF, SolarWinds Web Help Desk |
| Feb 10 | 6 | **Six Microsoft zero-days** from Patch Tuesday — Windows Shell, MSHTML, Office Word, RDP Services |
| Feb 13 | 1 | BeyondTrust CVE-2026-1731 (CVSS 9.9, ransomware flag) |
| Feb 20 | 2 | Roundcube deserialization + XSS |
| **Total** | **13** | Four batches in 20 days |
Thirteen known exploited vulnerabilities added to the federal catalog in three weeks. Six of them Microsoft zero-days. One with a ransomware flag. One CVSS 9.9 deserialization with 48-hour weaponization.
What Our Feed Shows (and Doesn't)
We ran every headline in this article against our IOC index — 894,879 indicators across 16,022 OTX pulses.
**FortiGate campaign IOCs**: Not in our index. Gap identified.
**BeyondTrust CVE-2026-1731**: Not in CISA KEV index. Gap identified.
**Roundcube CVEs**: Not in CISA KEV index. Gap identified.
We're ingesting all of these via our STIX 2.1 endpoint as this article publishes. By the time you read this, our 275+ STIX feed consumers in 46 countries will have these indicators in their next pull.
That's the point. The feed doesn't just aggregate — it responds. When we find gaps, we fill them. When CISA adds KEVs, we add KEVs. When Amazon drops a threat report, we ingest the IOCs.
275+ consumers. 46 countries. $0/month for the feed. The same data costs $50,000/year from the vendors who missed it.
*Subscribe to our STIX 2.1 feed at [analytics.dugganusa.com/api/v1/stix-feed](https://analytics.dugganusa.com/api/v1/stix-feed). 894,879 indicators. 350 named threat actors. All free. All legal.*
*DugganUSA LLC — protect. publish. amplify.*
*Her name was Renee Nicole Good.*
*His name was Alex Jeffery Pretti.*
Comments