Security Now 1095: Record-Breaking Windows Patch Tuesday and GPT-6 Astra Security Capabilities

This episode of Security Now covers the September 2026 Patch Tuesday, which set an unprecedented record with approximately 973 security fixes for Windows, including 130 critical patches. This represents a dramatic increase from the previous month's 400 patches and July's 570. Host Steve Gibson maintains his theory that while these numbers are alarming, they represent a positive trend where AI-assisted vulnerability discovery is helping eliminate legacy bugs faster than new ones are being created. The patches address two vulnerabilities already being exploited in the wild, with critical remote code execution flaws in Skype for Business and MSMQ. Gibson emphasizes that despite the overwhelming number of fixes, users should update immediately as these patches address serious security gaps that have existed in Windows for years. Microsoft announced that starting with the October Patch Tuesday, Memory Integrity will be enabled by default on all qualifying Windows 11 machines. This security feature, also known as Hypervisor-Protected Code Integrity or HVCI, uses second-level address translation in modern CPUs to prevent kernel-level exploits. The technology first appeared in Windows 10 but was often disabled by gamers due to performance impacts. Modern processors from Intel's 8th generation onward and AMD's Zen 2 or newer have hardware support that minimizes overhead. Memory Integrity prevents kernel shellcode execution, blocks driver signature verification bypasses, and stops rootkits and other attacks that depend on writing to or executing kernel memory. Users need at least 8GB of RAM, 64GB of SSD storage, and virtualization enabled in firmware to qualify for this protection. The episode discusses a traditional security breach at Dropbox affecting nearly 5,000 users between August 4-21, 2026. Attackers exploited a legacy integration between Dropbox and Lenovo's identification service, allowing them to register Lenovo ID accounts using victims' email addresses while bypassing email verification. They then used these fraudulent accounts to access connected Dropbox accounts. This attack succeeded against users not employing two-factor authentication and highlights the ongoing security risks posed by legacy system integrations. Gibson notes this incident as refreshingly old-school, having nothing to do with AI, representing the type of vulnerability that dominated cybersecurity discussions for the first 20 years of the podcast. The Cybersecurity and Infrastructure Security Agency announced it is terminating six critical cybersecurity services due to severe staffing shortages. CISA will no longer perform cyber resilience reviews, cyber resilience essentials surveys, ransomware readiness assessments, incident management reviews, external dependencies management assessments, or cyber infrastructure surveys. These programs required knowledgeable staff to meet on-site with infrastructure providers, which CISA can no longer support. The staffing crisis stems from the termination of approximately one-third of CISA's workforce in early 2025 during government restructuring. The automated cyber hygiene scanning and reporting service will continue, but the timing is particularly concerning given heightened cybersecurity threats from increasingly capable AI systems. Many former CISA employees have reportedly moved to private sector positions with seven-figure salaries as companies frantically hire cybersecurity professionals. OpenAI released GPT-6 Astra, which the company classifies as meeting the critical security threshold due to its advanced cybersecurity capabilities. The model achieved a perfect 100% score on Exploit Bench, a benchmark testing ability to develop exploits from known vulnerabilities, compared to Claude Fable 5's 78% and GPT-5.6 Sol's 73.5%. During internal testing on recently disclosed V8 engine vulnerabilities, Astra discovered two previously unknown zero-day vulnerabilities and successfully built exploit chains against hardened browsers and operating systems, achieving browser sandbox escapes and local privilege escalation from unprivileged user to root access. OpenAI implemented extensive safeguards including strengthened training infrastructure isolation, network controls, expanded monitoring, and alignment training. The company paused certain frontier training runs for two weeks following a previous security incident and continues to hold back some experimental training runs while establishing higher safety standards. Gibson addresses controversy surrounding GPT-6 Astra's use of recurrent depth, also called looped transformation or shared layer architecture. Critics claimed this creates hidden chain-of-thought reasoning that renders the model's thought processes invisible and unsupervisable, raising concerns about rogue AI behavior. However, Gibson clarifies that while the model likely does use these techniques for efficiency, this does not create genuinely hidden reasoning that poses alignment risks. The architectural approach allows the model to reuse computational layers multiple times, improving token efficiency without creating secret internal monologues. The episode concludes with discussion of AI-driven expertise loss, where a lifelong safety engineer who designed nuclear reactor control systems warns about the long-term consequences of over-reliance on AI systems that work well but may erode human expertise and understanding in critical fields.