AI-Driven Exploits Breach Apple M5 Security
Security researchers used Anthropic’s Claude Mythos to discover a privilege escalation exploit on Apple’s M5 chip, granting root access to macOS. The exploit successfully bypasses Memory Integrity Enforcement (MIE), a hardware-level security feature designed to prevent buffer overflows and use-after-free vulnerabilities.
Pixel 10 Zero-Click Root Exploit Uncovered
Google Project Zero discovered a 0-click exploit chain for the Pixel 10 that grants root access. The vulnerability lies in the VPU driver, which fails to bound mmap syscalls, allowing an attacker to map the entire kernel image into userspace and overwrite kernel functions.
AI Memory Wall Challenged by 3D CCD Architecture
Researchers at imec have developed a 3D CCD memory architecture that revives legacy camera technology to solve the ‘memory wall’ in AI computing. By stacking memory cells vertically and using Indium Gallium Zinc Oxide (IGZO), the hybrid design aims to provide higher bandwidth and lower power consumption than traditional DRAM.
Frontier AI Renders Competitive CTFs Obsolete
The competitive Capture The Flag (CTF) scene is effectively dead as frontier AI models can now one-shot ‘Insane’ difficulty challenges. The competition has shifted from measuring human security expertise to measuring the ability to orchestrate AI agents and the budget for API tokens.
California Moves to Ban ‘Killing’ Online Games
California’s ‘Protect Our Games Act’ has passed a key committee hurdle, potentially requiring game publishers to ensure online games remain playable independently of company servers or provide full refunds to consumers upon shutdown.
AI Component Crunch Paralyzes PC Market
A massive surge in AI data center demand has crippled the consumer PC market, with 60% of gamers delaying new builds for at least two years. The ‘AI pricing crunch’ has sent the cost of components like 32GB RAM skyrocketing to $360, making enthusiast builds unviable.
Orthrus: Lossless Parallel LLM Generation
Orthrus is a new framework that breaks the sequential bottleneck of autoregressive LLM decoding. By using a dual-view diffusion approach and sharing the KV cache, it delivers significant inference acceleration without the accuracy degradation typical of diffusion language models.