AI news, 6 November 2026

OpenAI ships GPT-6 Cyber for authorised security research (Daybreak Red)

OpenAI launched GPT-6 Cyber, a cybersecurity-specialised model it says finds and validates exploitable bugs while staying inside researcher-set scope. OpenAI classifies it as Critical under its Preparedness Framework (up from High for GPT-5.6 Cyber), reports 48.5% on ExploitGym versus 32.1% for the prior cyber model, and says the same stack powered the Linux research (948 working local privilege-escalation exploits) plus 81 previously unknown high-severity Firefox findings Mozilla has started fixing. Access starts with select Daybreak Red API and Codex customers. Benchmarks and customer quotes are OpenAI's. Fact or hype: fact that the product page, system card and Daybreak Red availability exist; partly fact on every capability percentage — those are OpenAI's evals and partner quotes.

Why it mattersIf you are a vetted security team, apply for Daybreak and treat it as a research tool inside a sandbox with human review. Everyone else: keep patching Linux and browser stacks on a schedule.

Sourcesx.com post ↗ · x.com post ↗ · openai.com ↗Partly fact Hype or Fact? check

• ChatGPT (OpenAI)
• OpenAI Developers (OpenAI Platform & Codex)

Thinking Machines' Inkling 2 Small: big MoE on Tinker, weights next week

Thinking Machines Lab released Inkling 2 Small: a mixture-of-experts model with 594B total parameters and 17B active, image inputs, and a score of 31 on the Artificial Analysis Intelligence Index (company/third-party reported). It is available now for fine-tuning and inference on Tinker; full weights are planned under Apache 2.0 after external safety testing. Fact or hype: partly fact. Tinker access and the announced licence path are real; leaderboard wins are vendor/third-party snapshots until you run your own tasks.

Why it mattersAnother US open-weight candidate aimed at agent loops and fine-tuning. Try it on Tinker for a real workflow this week; decide on self-hosting only after the weights land.

Sourcesx.com post ↗ · x.com post ↗ · thinkingmachines.ai ↗Partly fact Hype or Fact? check

• Thinking Machines Lab

METR: Anthropic's AI progress ~1.9× faster, with AI helping build AI

Independent evaluator METR published an assessment of AI R&D acceleration inside Anthropic. It estimates that since October 2025, Anthropic's AECI benchmark scores have grown at about 1.9× the earlier rate (95% interval roughly 1.5–2.8), and that AI itself may account for a large share of that speed-up — METR flags that share as highly uncertain. Anthropic redacted much of the non-public evidence in the public report; METR still judges catastrophic automated-R&D risk as low for models as of mid-July. Fact or hype: partly fact. The public report exists; the multipliers are METR's with wide uncertainty and heavy redactions.

Why it mattersLab roadmaps may compress. Don't rewrite your tool stack on a single multiplier — do budget for more frequent model swaps and re-tests of agent workflows.

Sourcesx.com post ↗ · x.com post ↗ · metr.org ↗Partly fact Hype or Fact? check

• Claude & Claude Code (Anthropic)

Matt Pocock's one-line verification prompt for coding agents

@mattpocockuk's prompt of the day: "How do we make this change fully verifiable?" He says it pushes coding agents to sign up for real services and click through the UI instead of declaring victory. About 1.7k bookmarks at read. Opinion/tip, not a product claim.

Why it mattersA free habit for any Claude Code / Codex / Cursor loop. Add it to your project instructions when agents ship UI or integrations.

Sourcesx.com post ↗

• Claude & Claude Code (Anthropic)

AI agent observability map for regulated workflows

@MaryamMiradi posted a photo roundup of observability tools for agents in regulated work (medical, finance, legal, cyber): knowing which tool failed, why a step retried, and which hop blew the budget. Names include Langfuse, Phoenix, Opik, AgentOps, OpenLIT and Helicone. Useful as a shopping list, not a hands-on review.

Why it mattersIf an agent touches customer data or money, logging and trace IDs matter as much as the model. Pick one tracer before you scale seat count.

Sourcesx.com post ↗

The Second Brain Guide: Claude + Obsidian personal wiki

An X Article, "The Second Brain Guide," covers a Karpathy-style personal wiki: Claude plus Obsidian, with a brain.py linter so notes stay structured. Practical build framing rather than a SaaS pitch. Saved as a bookmark.

Why it mattersA concrete pattern for stopping re-explain-yourself loops with coding agents. Treat it as a weekend project, not a vendor claim.

Sourcesx.com post ↗ · x.com post ↗

• Obsidian