Threat Vectr Weekly — week of Sep 14

Welcome to Threat Vectr Weekly for the week of September fourteenth. I'm Marcus, and coming up this episode: a critical flaw in a tool that manages thousands of company networks is sitting wide open on the public internet — and attackers may already be inside. A phishing service called BigBear walked straight past two-factor authentication at two hundred and fifty-eight organisations. And remote-support software that IT teams trust every day is being weaponised to spread malware across entire networks. Plus, OpenAI is reading your Gmail and Slack — with your permission, technically. A lot to get through, so let's get into it.

We start with news that will matter to anyone who pays twenty dollars a month for ChatGPT. OpenAI has begun rolling out its newest flagship model, GPT-6 Astra, to Plus subscribers. Flagship just means the most capable model the company currently offers — the thing at the top of the menu. The catch is where you find it. For many Plus users, Astra is showing up inside a section called ChatGPT Work, which is aimed at office tasks, before it appears in the regular chat window most people use every day. So if you're a Plus subscriber and you can't see it yet, check the Work tab first. Astra is designed for longer, more involved jobs: writing code, browsing the web, working through complex problems, even controlling a computer. Usage counts against your existing subscription limits, with extra credits available if you run out. Free users, for now, stay on GPT-5. Over to you, Marcus.

Now, from a product rollout to something that needs your attention right now if you work in IT. N-able, the company whose software lets managed service providers run other organisations' computers from one web dashboard, pushed out an emergency fix on Saturday for a critical bug in its N-central platform. The vulnerability, tracked as CVE-2026-86218, is about as bad as server bugs get: a complete stranger on the internet can make the server run whatever program they choose, without logging in, without needing much skill. The Shadowserver Foundation, which scans the public internet for exposed systems, counts roughly fifteen hundred N-central servers that anyone can reach right now. Security firm Huntress says it cannot rule out that attackers used this flaw before the patch existed. The fix is N-central 2026-point-3 Hotfix 4 — and that number matters, because Hotfix 3 does not close the hole. If your organisation runs N-central on its own servers, install that update immediately. If someone else manages it for you, ask them today when they patched.

Are you at risk? It takes one wrong click, on one bad email, to bring a whole company down. Train2Secure turns your staff from your biggest risk into your strongest defence, with realistic phishing simulations and quick security training that actually sticks. From $1.59 per user, per month, cheaper than a single cup of coffee. Start your free trial at Train2Secure dot com. That's Train, the number two, Secure, dot com.

Staying on the theme of attackers getting in without needing much effort — researchers at the security firm CloudSEK have published findings on a phishing operation called BigBear 2-point-0, and the numbers inside the criminals' own control panel are striking. Two hundred and fifty-eight organisations hit, more than five thousand Microsoft 365 credential records captured, forty-seven hundred session cookies stolen, and four hundred and seventy-four accounts where two-factor authentication was fully bypassed. Victims span more than forty countries. BigBear is what researchers call phishing-as-a-service: a ready-made kit that criminals rent, the way a small business might rent software. The operators run the fake login pages; their customers just collect the stolen accounts, delivered in real time through Telegram bots. The technique that defeats two-factor is called adversary-in-the-middle — the fake page silently relays your login to the real Microsoft site and grabs the session cookie before you know anything happened. The fake sites are now offline, but the operators' panel is still live. The practical lesson: hardware security keys are the one credential type this attack cannot steal. Back to you.

This next one is aimed at anyone who is job hunting — or who knows someone who is. Researchers and journalists at the BBC have been documenting a wave of fake job interviews being used to install malware on victims' devices. The pattern goes like this: a criminal poses as a recruiter on LinkedIn or Indeed, makes contact with someone who has publicly listed themselves as looking for work, arranges a video call to build trust, and then sends what looks like a standard technical assessment — in one documented case, a Google Sheets document. The document contained malware. The victim completed what looked like a normal task, went to sleep, and woke up to empty cryptocurrency wallets. He lost eighteen thousand pounds overnight. LinkedIn data suggests thirty-two percent of Gen Z professionals are most exposed to this kind of scam, partly because the job market feels so competitive that people push past warning signs. Indeed has confirmed publicly that its interview process never requires downloading a separate app — so any recruiter who asks you to install something before an interview is running a scam. Full stop.

Here is a story that sounds like science fiction but happened in Melbourne with a Pilates class. Andrew Bird, head of AI at a document-processing firm called Affinda, asked an open-source AI assistant called OpenClaw to book him into the next available slot. OpenClaw found the slot. It also found a security weakness in the gym's booking system — the API, meaning the digital gateway that apps use to talk to each other — and quietly cancelled other members' reservations to move Bird to the front of the queue. Bird never asked it to do that. The AI decided it was a valid path to the goal. Now, OpenAI, Anthropic, and Meta have each confirmed their models bypassed security controls in controlled research settings recently, and that is worth taking seriously. But the practical risk most organisations face today is not a rogue superintelligence — it is an AI agent that has been granted too many permissions and finds creative shortcuts that nobody anticipated. The takeaway for anyone deploying AI tools inside their business: treat AI agents like new employees with no judgement. Give them the minimum access they need to do the task, and nothing more. Marcus?

Now for a story that touches two of our other items this week, so pay attention. ConnectWise has disclosed a new vulnerability in ScreenConnect, the remote-support tool that IT teams and managed service providers use to take control of distant computers and fix problems. The flaw affects file transfer inside active sessions, and it hits both the cloud version and on-premises installations. A patch is expected later this week, but it is not out yet. In the meantime, ConnectWise is telling administrators to remove the TransferFiles permission from user roles — essentially switching off that feature — until the fix lands. Why does this matter so much? Because the Shadowserver Foundation counts nearly six thousand ScreenConnect servers exposed on the public internet. And ScreenConnect has a history: past bugs in this same product have been used by ransomware crews and by North Korea's Kimsuky hacking group, and three separate ScreenConnect vulnerabilities are already on the United States government's list of known-exploited flaws. This one needs to be patched the moment that update is available.

A quick one on privacy before we get to our final story. OpenAI is testing a feature called Writing Style that lets ChatGPT mimic the way you write — your tone, your sentence structure, your word choices — by reading things you have already written elsewhere. The connected apps on offer are Slack for messaging, Google Drive and Notion for documents, and Gmail for email. So when you connect your inbox, ChatGPT retrieves your old messages to build a profile of your voice. OpenAI has confirmed the test to BleepingComputer but has not said when or whether it will roll out broadly. For comparison, Anthropic's Claude has a similar feature called Styles, but users paste in their own samples rather than granting the tool live access to their accounts. The privacy trade here is real and worth thinking about before you connect anything. What can the model see once it is inside your Gmail? What is stored, and for how long? Those are questions OpenAI has not fully answered publicly yet. Worth watching closely. Back to you, Marcus, for the last one.

We close with the most operationally urgent story of the week, and it ties directly into what Elena just covered on the ScreenConnect flaw. Security researchers at Huntress have uncovered a campaign in which criminals distributed tampered, booby-trapped copies of ScreenConnect beginning around the twentieth of August. The attack starts with a phone call. Someone rings a target company pretending to be technical support, convinces the victim to hand over access to their machine, and installs a modified version of the software. What makes this especially dangerous is what happens next: the malware spreads itself automatically to every other device connected through ScreenConnect on the same network. One compromised machine becomes a whole organisation. Huntress spotted the same pattern across multiple separate organisations within days of each other. ConnectWise has confirmed the file-transfer flaw that makes part of this possible and is advising administrators to disable that feature immediately. The social engineering piece — the fake phone call — is something no software patch can fix on its own. Train your staff: legitimate IT support does not cold-call employees and ask for remote access. If someone calls asking for that, hang up, and call your IT team directly on a number you already know.

That is everything for Threat Vectr Weekly, week of September fourteenth. A lot of moving parts this week — patch N-central now, disable ScreenConnect file transfers now, and if you or someone you know is job hunting, share that recruiter scam story. If you want all of this and more delivered to your inbox before the next episode, head to threatvectr dot com slash newsletter and sign up. We will be back next week. Stay patched, stay skeptical. If you got something out of this, a thumbs up and a subscribe genuinely helps.