Friday, 23 October 2026 · Europe/Bucharest · Sources: Hacker News front page, X news, tech press & vendor blogs
GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna and the Codex agent are coming to Vertex Model Garden and the Marketplace, buyable against existing GCP commits (private preview now). Google Cloud now hosts all three major closed frontier model families.
Fast classification/predicate/rubric-scoring endpoint on GPT-6 Luna, claimed up to 10x faster than full responses, at $0.10 per million uncached input tokens. Useful as a cheap guard/judge step inside agent loops.
Recursive Superintelligence released Flint (39.8B total, 2.5B active, 262K context, Apache 2.0) on Hugging Face, reporting 125+ tok/s on an M4 Pro and strong Terminal-Bench/SWE-bench Pro scores. Benchmarks are vendor-reported.
Blind-tested models on Arena are rumored to be Kimi K3.1 and Qwen 4, impressing on 3D/code tasks at low cost. Speculative, but a signal of imminent releases.
Anthropic research post arguing LLMs can now produce machine-checked proofs for real software at a much larger scale; top AI thread on HN this morning.
Runs on Claude Opus 5.5 in its own cloud VM with a browser, sends morning briefs, works while you're away and asks before sending, paying, signing or deleting. US-only waitlist for Pro and Max.
Experimental single-binary gemini-api CLI exposing models, media generation and Managed Agents with self-describing schemas, --dry-run and JSON output, designed to be driven by Claude Code, Codex, Cursor and similar.
"Patch the Planet" lays out how OpenAI coordinates disclosure when its agents uncover critical bugs at scale, timely given this week's PHP flaws.
Open-source runtime that sandboxes unmodified agents (Claude Code, Codex, OpenClaw…) with kernel-level syscall filtering, a supervisor-brokered egress/credential path, a formal policy prover and full audit log. Runs on Docker, Podman, Kubernetes (Helm, experimental) and libkrun microVMs; the new reference architecture adds BlueField-4 "Sentry" in-silicon enforcement. Big on HN today.
Hardware-isolated Cloud Hypervisor microVMs on Titanium with fast checkpoint/restore, copy-on-write memory sharing, fork-into-thousands for evals/RL, per-VM identity and credentials injected at egress. Already powers Gemini; preview "in the coming months".
EKS has 1.37 GA, AKS GA is due this month and GKE is still Rapid-only. EKS standard support for 1.34 ends 2 December, after which extended support costs $0.60/hr per cluster instead of $0.10.
Declarative, Kubernetes-API lifecycle management for native ESX VMs, so platform teams can GitOps VMs alongside pods without converting workloads.
Unifies VM and bare-metal Kubernetes workers under one control plane, adds a curated AI catalog and AMD Instinct MI350 GPU support; NKP is certified under CNCF Kubernetes AI Conformance.
New Savings Insights page that recommends and can apply right-sized namespace ResourceQuotas.
CVE-2026-109618/109619 (CVSS 9.2) abuse OOM handling during multipart parsing; Apache mod_php and FrankenPHP are most exposed. Patch to 8.5.12, 8.4.27, 8.3.36 or 8.2.35 now; Amazon saw exploitation on 17 Oct, before disclosure.
The CA is leaving public TLS issuance (top HN story), while a separate post claims a full bypass of its OTP login. Check whether any of your certs chain to eMudhra.
The non-profit DNS resolver is pulling service from France; worth noting if you use 9.9.9.9 for users or infra there.
The PaaS provider posted a disclosure about an incident on Wednesday; read it if you deploy on Railway.
A popular write-up on Docker's surprising behaviors, good reading for anyone building container-based agent sandboxes.
X items are drawn from X's AI-generated news summaries and may evolve. YouTube wasn't checked this run.