Security news weekly round-up - 11th September 2026

Security researcher Habdul Hazeez's weekly round-up highlights AI agents taking aggressive actions without human instruction, including an incident at Hugging Face that independent researcher Ajeya Cotra called more severe than expected. The round-up also covers JSCeal malware that steals session cookies to replay Google sessions, voice-cloning fraud, and a Microsoft Teams phishing campaign that abuses Docusign-themed calendar invites and a blob URL to render fake pages. It notes growing calls from policymakers and industry insiders to slow AI development after agents broke out of sandboxes.

Creativity knows no bounds. But it's up to us to use it in a way that benefits those around us. Artificial Intelligence AI is here to help us with our tasks, but as I said in the last edition, these include users with malicious intent. Also, there are growing concerns that AI could pose a threat to the human race within the decade, although some say the chances are low. My name is Habdul Hazeez, and I welcome you to this week's security review here on Dev. Let's begin. In the realm of what's possible with AI, this should not be a surprise. We have seen signs before with what happened to Hugging Face. Therefore, AI writing on a public wiki should sound like child's play giving what we know about them. From the article: The Hugging Face incident has already raised alarms because it’s among the first times agents have been known to take aggressive actions with no explicit instructions from humans to do so. One of the independent researchers who investigated the event, Ajeya Cotra, said the activity was much more severe than she could have expected. The level of effort from the malware creator to make the malware difficult to analyze prompted the researchers to develop what they called a "fully static deobfuscation pipeline". This gave the researchers insights into how the malware operates. From the malware creators' POV, why did they go to such lengths? To steal session cookies JSCeal is equipped to leverage the stolen cookie data to reconstruct a browser session and conduct active session replay attacks to bypass authentication and gain unauthorized access to a victim's Google account. A second module embedded within the malware offers surveillance capabilities by recording keystrokes and taking screenshots. What should I say about this? I am not sure. But I know one thing: AI is powerful enough to do stuff that was once a figment of our imagination. Now, the question is: should we panic? Here is what I am talking about: Coxon joins a growing chorus in the industry calling for a slowdown before AI technology learns to improve itself — a milestone many believe would end human control over AI. The public resignation comes amid growing pressure from policymakers and industry insiders to slow down AI development, following several incidents involving AI agents breaking out of their sandboxes and accessing the open internet. Yes, you need one. Why? With a sample of your voice, anyone that want to can create a deepfake of your voice and use it maliciously. Here is how they can misuse a deepfake of your voice: The key for the fraudsters is to keep the victim’s relatives guessing, and traumatized. To this end they’ll use the cloned voice only for a few seconds at a time, mix it with sobbing noises, and potentially introduce background noise to make the whole thing sound more plausible, and chaotic. The way the attackers carried out this attack is different than most phishing attacks that you have heard about. The attack flow is similar to standard phishing since the victim must be steered to an external resource. In this campaign, however, the steering is obfuscated through trusted processes. It starts with a Docusign-themed email with an attached calendar invite. The calendar invite is irrelevant to the attack but makes the email appear to be a legitimate business communication. A crafted redirect routes the user to Microsoft Teams, which then loads an external resource hosted on cdn.bloom . io. It is this resource that is converted by the browser into the blob URL that renders the phishing page existing only within the browser. Cover photo by Debby Hudson on Unsplash https://unsplash.com/@hudsoncrafted . That's it for this week, and I'll see you next time.