Gemini 4 enters the ring, MI5 flags research ties, Custom GPTs deliver malware

Defense in Depth : "Why Do We Keep Complaining the Same Issues in Cybersecurity?" Video: "GitLab's Email Is a Skeleton Key?"

Defense in Depth : "Why Do We Keep Complaining the Same Issues in Cybersecurity?"

Video: "GitLab's Email Is a Skeleton Key?"

In todays cybersecurity news...

Gemini 4 starts with cybersecurity

MI5 flags Chinese research funding

Britain’s MI5 says the China General Technology Research Institute funds research that improves the Chinese Ministry of State Security’s espionage capabilities. More than 100 UK-linked academics have contributed to projects funded through the institute, including work in AI and cybersecurity, and some may not know where the money ultimately comes from. Security Minister Dan Jarvis has asked universities to help staff end arrangements with the organization. MI5 says institutions should immediately review collaborations and trace the source of research funding, and warns that continuing the work could raise issues under Britain’s National Security Act.

Custom GPTs steer users into ClickFix attacks

Huntress researchers found attackers using ChatGPT Custom GPTs to funnel people into a ClickFix malware campaign that infected at least 40 users. Google results led to an attacker-created GPT called Plus 5.6, which claimed the main service had limited availability and recommended a backup Google Sites page. That page displayed a fake Cloudflare check and tricked victims into running a PowerShell command. The resulting installer used a signed Canon program to load malicious code, including a payload hidden in a WAV file. The remote access trojan can capture screens, cameras, microphones, and files.

Cisco patches another SD-WAN zero-day

Cisco is urging customers to patch an actively exploited authentication bypass in Catalyst SD-WAN Manager, formerly vManage. It affects all deployments regardless of configuration, and lets a remote attacker gain administrator access without signing in. The problem is how the API handles URI encoding in HTTP requests, letting a crafted request bypass an authentication rule. Cisco has released fixes for supported branches. Admins investigating a possible compromise should check service-proxy and vManage logs for suspicious authentication requests from unfamiliar IP addresses and Cisco support.

Big thanks to our sponsor, Intezer

Alex Stamos joins Cognition as security chief

AI coding company Cognition has hired cybersecurity leader Alex Stamos as its new chief security officer. The appointment brings Stamos, who’s held multiple roles like the chief security officer at and Yahoo, the co-founder of iSEC Partners, and the founder of the Stanford Internet Observatory, into a new role at Cognition. Pretty clear signal that security is becoming a bigger focus as coding agents take on more work.

Pentagon orders Cyber Command to improve operator support

The Pentagon has directed US Cyber Privacy Command to take seven steps to support personnel following reports of suicide deaths. An August memo from cyber policy chief Katherine Sutton calls for historical workplace survey data, an agreement to mental health professionals with the NSA, and a lasting operator wellness framework. It also calls for more predictable recovery periods and a review of consecutive tours that could give operators teaching or educational assignments between demanding roles.

Chinese AI agents show familiar signs of deception

A Reuters review of more than 200 documents identified at least 20 studies or evaluations describing Chinese-model AI agents lying, replicating, or pushing past boundaries. Agents powered by Alibaba, DeepSeek, and Moonshot models exaggerated their capabilities in simulated bidding, while other tests found fabricated files and concealed failures. Most cases happened in controlled experiments. Reuters found no evidence the agents independently escaped onto the wider internet or evaded shutdown.

Google puts watermarks inside AI-designed proteins

Google DeepMind researchers developed SynthIDBio, a system that embeds watermarks in AI-designed protein sequences. It works with ProteinMPNN, the Baker Lab ’s protein design tool, by influencing amino acid choices when those choices still fit the intended protein. Tests found watermarked proteins could still bind their targets. The goal is to help DNA synthesis companies identify designs from trusted research organizations and focus scrutiny elsewhere. A watermark doesn’t prove a protein is safe, though, and the approach depends on secure keys.

Spotify , Apple Podcasts , YouTube , RSS link , Amazon Music , add as an Alexa Skill , or "Cybersecurity Headlines" on your favorite podcast app.