AI Servers Under Siege, ShinyHunters Leader Caught & Two Perfect-10 Flaws
Wednesday, October 7, 2026 · 5-minute read
A teenager in Amman, Jordan — known online as "Rey" and identified as Saif Al-din Khader — has been detained by Jordanian authorities. He's suspected of leading ShinyHunters, one of the most active data theft and extortion gangs online. Khader is reportedly cooperating with the FBI to identify other members of the group.
His arrest came while ShinyHunters was actively trying to extort Jeppesen ForeFlight, a aviation software firm recently sold by Boeing. Sources told Krebs on Security the stolen data could pose real safety risks to aviation operations.
↗ Krebs on SecuritySouth Korean authorities are investigating a string of bank breaches they believe may have been carried out using AI-powered attacks. The specifics of how the attacks worked haven't been publicly confirmed yet, but officials are treating the incidents as serious. It's another sign that criminals are starting to bake AI into their hacking toolkit.
↗ BleepingComputerUK online fashion giant ASOS confirmed a data breach after hackers sent unauthorized push notifications through its mobile app — literally announcing the hack to customers in real time. The attackers claim to have stolen customer data from ASOS's Snowflake environment. If you shop at ASOS, watch your email for a breach notification and consider changing your password.
↗ BleepingComputerThe FBI and Secret Service issued a fresh warning Tuesday: a campaign called FortiBleed is still actively hitting internet-facing Fortinet firewalls. Attackers — believed to be a Russian-speaking group — collected over 86,644 working login credentials from Fortinet devices across 194 countries. They're now using those stolen credentials to keep breaking in, while also cracking older password formats. The credentials were first documented back in June 2026, but the attacks haven't stopped.
Hackers are actively exploiting a critical bug (CVE-2026-21589) in Atlassian's popular collaboration tools — Jira, Confluence, and Bitbucket — without needing a password. The flaw allows an outsider to access files they should never be able to reach. Atlassian disclosed the vulnerability just days ago, but attackers jumped on it as soon as a working proof-of-concept was made public. Companies running self-hosted "Data Center" versions are the ones at risk.
SonicWall's SMA1000 appliances — the gateways that let remote workers connect to a company's internal network — have a perfect-10 severity flaw. An attacker who isn't logged in at all can send specially crafted requests through the device's WorkPlace login portal, reach internal systems, and perform unauthorized actions. Think of it like a secret side door into a building that bypasses the front desk entirely. Three SMA1000 models are affected: the 6210, 7210, and 8200v.
Status: Hotfix patches available from SonicWall. No active exploitation reported yet, but a perfect 10.0 score means you should patch immediately.
LMCache is open-source software that speeds up LLM servers like vLLM. A critical flaw lets anyone on the network send a single message to the LMCache server and run any command they want — no login required. The server is only reachable from outside when set up to listen on a public address rather than staying local, but many AI infrastructure operators may not realize theirs is exposed. Versions 0.3.9 through 0.5.5 are all affected.
Status: No fixed version available yet. If you run LMCache in multiprocess mode with a public-facing address, restrict network access immediately while waiting for a patch.
This flaw lets an unauthenticated attacker read files on self-hosted Atlassian servers that they should have no access to. That could include configuration files, credentials, or sensitive project data. It affects Atlassian's Data Center products — the self-managed versions companies run on their own servers. Within days of a public proof-of-concept being released, hackers were already using it in real attacks.
Status: Patch available from Atlassian. Apply it immediately — exploitation is confirmed and ongoing.
OpenAI is rolling out invisible watermarks on text produced by ChatGPT and its Codex coding tool for users in the European Union. The watermarks are designed to let platforms and researchers identify AI-generated content without the text looking any different to a human reader. This is part of a broader push — driven partly by EU regulations — to make it easier to detect AI-written articles, emails, and code. It's an early but significant step toward accountability in AI-generated content, and other AI providers may follow suit under similar regulatory pressure.
Starting next month, Microsoft will block .msix and .msixbundle attachments in Outlook Web and the new Outlook for Windows. Hackers have increasingly used this file format to deliver malware because it can bypass some security filters. Microsoft's move follows a pattern of the company adding dangerous file types to Outlook's block list. If you or your colleagues regularly share .msix files via email, you'll need to find an alternative delivery method — like a shared drive — before November.
A new malware campaign called Canto Incognito has infected more than 3,400 servers — and the targets aren't banks or hospitals. They're AI servers. Specifically, exposed infrastructure running large language models. Researchers at Lumen Black Lotus Labs say the campaign is installing cryptomining software and turning infected machines into launching pads to find and attack even more vulnerable servers.
Here's how it works. Companies and researchers who run their own AI infrastructure — the servers that power LLMs — sometimes leave management ports and APIs exposed to the open internet without realizing it. The PoeLLM malware scans for these exposed systems, breaks in, and installs crypto miners like XMRig and Iron. It then connects the compromised server to Kryptex, a Russian cryptocurrency mining service, and immediately starts using that server to hunt for the next victim. The malware's name comes from a clever trick: it hides the address of its command center inside a poem that the attackers wrote — a genuinely novel technique researchers called "creative."
Why does this matter to people who aren't running AI servers? Two reasons. First, if your company is exploring AI infrastructure — even a small internal LLM deployment — this is a concrete warning that poorly configured AI servers are actively being hunted. Second, the campaign demonstrates that as AI infrastructure spreads, attackers are spreading with it. The same gold-rush mentality that drove companies to spin up AI servers quickly is creating exactly the kind of exposure that campaigns like Canto Incognito feed on. Speed and security don't always travel together.
The broader pattern here is worth watching. This is the second major AI infrastructure vulnerability in today's newsletter alone — the LMCache flaw also affects LLM servers. Attackers are clearly paying attention to where the new computing power is concentrated. If AI servers become as common as web servers, they'll need to be secured just as carefully. Right now, many aren't.