OpenAI’s GPT-5.5 is as Good as Mythos at Finding Security Vulnerabilities

This feed aggregates multiple high‑impact security stories: the UK AI Security Institute found OpenAI’s GPT-5.5 comparable to Anthropic’s Claude Mythos at finding vulnerabilities (models increasingly effective offensive tools). TheCopy.Fail Linux kernel local privilege escalation (CVE-2026-31431) abuses AF_ALG sockets + splice() to write into another file’s page cache, works unmodified across major distributions, and evades checksum/integrity detectors. New GPU rowhammer research demonstrates GDDR bitflips on NVIDIA Ampere cards that can lead to full host compromise when IOMMU is disabled. GTI