Security firms have discovered a new attack vector where malicious custom GPTs hosted on ChatGPT's official platform are used to infect Windows users with remote-access malware. Victims are lured via search ads to fake authentication pages and tricked into executing malicious PowerShell commands. The attackers employ techniques such as Cloudflare spoofing, DLL sideloading using legitimate applications, and audio steganography to evade antivirus detection before deploying the final remote-control tool. Although the initial malicious GPT was removed, similar variants and abuses of other AI sharing features continue to be discovered, prompting warnings to monitor for suspicious script executions and app behaviors.

Chain-of-thought monitoring as an approach for AI oversight and control is threatened by the possibility of steganographic reasoning, where LLMs conceal their reasoning inside innocuous-looking text. Two neighbouring capabilities, steganographic messaging (passing a concealed message) and encoded r…

Archived notices for a CGTRI joint fund, reviewed by a research group, show priorities that moved from cryptography and communications security to AI agents, password

Chain-of-thought monitoring as an approach for AI oversight and control is threatened by the possibility of steganographic reasoning, where LLMs conceal their reasoning inside innocuous-looking text. ...

MI5 warned that Chinese spies targeted over 100 UK-linked academics, funding research into AI and steganography via a front for Beijing's intelligence agency.

Security service says academics should stop taking grants from and working with firm backed by China’s intelligence service

This September 22nd, 2026 SANS Internet Storm Center Stormcast covers three cybersecurity topics. The first discusses a PNG file used in a terminal fix campaign that Microsoft reported, where an executable was encoded inside the PNG pixel data using steganography. Unlike typical steganography, the entire image data was the executable itself, and the PE data was obfuscated by encoding differences in pixel values, requiring a Python tool called PNGdump.py to extract. The second topic covers a supply chain attack against npm called the B3 malware campaign, which uses a library named indexed B3 to mimic a legitimate sorted B3 library. This attack differs from previous npm attacks by avoiding install scripts and instead executing malicious code as part of the library's own code, evading current detection techniques and remaining on GitHub at the time of reporting. The third topic announces four vulnerabilities in Pi Hole, including two remote code execution vulnerabilities that require authenticated sessions to exploit.