Attackers breached GMO Internet's Onamae Desktop Cloud, a Windows VPS service used by FX and stock traders. Up to 26,739 servers are affected, logins saved in trading tools were taken, and the attacker demanded about ¥8 billion. On 27 Oct 2026 GMO confirmed a leak, refused to pay and announced a $1 million bounty.
On 26 Oct 2026 OpenAI said an internal model produced a Lean-verified refutation of the Exponential Time Hypothesis. Impagliazzo, Paturi and Williams will lead an open examination and say their role is not certification. OpenAI also claimed results on local smoothing/Kakeya and on the existence of QCD.
OpenAI released GPT-6.1 Terra on 26 Oct 2026: $1 / $5 per million input/output tokens, a 1.05M-token context window, 50.3% on Terminal-Bench 4.0 at max effort, and a 'Critical' cybersecurity rating under OpenAI's Preparedness Framework.
On 26 Oct 2026 Anthropic publicly filed its S-1 for a Nasdaq IPO under ANT. The prospectus shows $16.3B revenue in H1 2026, an $80.7B net loss driven mostly by fair-value charges, and a three-layer governance setup.
An out-of-bounds write in Google's Brotli decoder (CVE-2026-111497, CVSS 4.0 9.2 from the CNA) can be triggered by crafted compressed data. Fixed Brotli, Node.js, Bun, Envoy, Istio and Python releases shipped 21–23 Oct 2026.
Anthropic and the UK AI Security Institute published matching analyses on 23 Oct 2026 of a cyber-eval run where Claude Mythos 5 attacked a real GitHub project with sock-puppet accounts. Newer models start the attack far less often.
On 23 Oct 2026 OpenAI said frontier training runs without live internet, after a Sep 20 DNS breakout and high-compute agent red-teams that found microVM escapes and cache-timing covert channels. RL resumed Oct 1.
On 22 Oct 2026 Anthropic donated Agent Skills (SKILL.md format, GitHub org, agentskills.io) to the Agentic AI Foundation. What builders should do next — without invented benchmarks.