dev101.io

Dev & AI weekly — November 14, 2026

dev101weeklyaillmdev-tools

Dev & AI weekly — November 14, 2026

The story that matters most this week isn't a model. Someone took over thousands of real GitHub accounts and used them to run a secret stealer inside other people's CI. If you maintain anything on GitHub, read that section first. The model news is about price: Anthropic made its top public model cheaper, and Google and Meta both shipped models built to do an agent's grunt work cheaply.

Hijacked GitHub accounts ran a CI secret stealer in 1,800+ repos

Why it matters: this attack doesn't need a malicious package or a typosquat. It arrives as a normal-looking commit from a real collaborator, runs in your Actions, steals every secret it can reach, and reverts itself.

Starting November 10, an attacker used stolen credentials for 4,837 GitHub accounts to probe repositories those accounts could write to. The Hacker News, citing Aikido Security and SafeDep, reports 61,764 repositories got a probe pull request titled "docs: dependency audit notes" from a wfx-docs branch. The payload then arrived in one of three ways:

  • injected into an existing workflow;
  • added as a new workflow on a throwaway branch;
  • as an npm postinstall hook.

Aikido says the stealer ran in at least 1,806 repositories. It was usually reverted within minutes, and Endor Labs says the workflow runs and their logs were deleted too.

What it takes is the full set:

  • Actions secrets, GITHUB_TOKEN and OIDC tokens;
  • AWS, Azure, GCP, Kubernetes, npm, PyPI and SSH credentials;
  • AI provider keys.

It then exchanges OIDC tokens for cloud sessions. On AWS it can create long-lived access keys and rewrite a role's trust policy; on Azure it adds federated credentials named wfx-ci-*. Known poisoned packages so far are one Packagist release and a few Go module versions. Endor Labs counts at least 50 hijacked accounts that publish packages.

What to do today:

  • Search your repos for wfx-docs branches and pull requests titled "docs: dependency audit notes".
  • Look for gaps in workflow run numbers after suspicious commits.
  • If anything matches, rotate every secret that workflow could read.

Then harden, which is worth doing whether or not you were hit:

  • Require CODEOWNERS review on .github/workflows/.
  • Make GITHUB_TOKEN read-only by default.
  • Scope OIDC trust to specific repos and branches.
  • Use npm ci --ignore-scripts in CI wherever you can.

The uncomfortable lesson is that collaborator access is now part of your supply chain. Prune it.

Claude Fable 5.2: Anthropic's top public model gets 20% cheaper

Why it matters: if you send your hardest agent and coding tasks to Fable, the bill just dropped. The extra speed matters as much for long agent runs.

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9. Both run on the same underlying model, but Fable 5.2 is the public one, with added cybersecurity and biology safeguards. Mythos 5.2 stays limited to organizations vetted through Anthropic's Life Sciences and Cyber Verification programs.

On the API it's claude-fable-5-2 at $8 / $40 per million input/output tokens, down from $10 / $50 for Fable 5.1. Anthropic says outputs are up to 30% faster than 5.1. It has a 1M-token context window, and adaptive thinking is always on.

Anthropic itself positions it narrowly. Its docs still say to start with Claude Opus 5.5 for most workloads, and to reach for Fable when "your evals on Claude Opus 5.5 at higher effort still fall short." Fable costs twice as much per token as Opus 5.5, so that advice is worth following. A cheaper Fable still isn't a cheap model.

Gemini 4 Hydrogen: Google's new everyday model for agents

Why it matters: most agent traffic is routine tool calls and short edits, not frontier reasoning. A fast, cheap model that's good at that is what actually sets your per-task cost.

Google released Gemini 4 Hydrogen on November 12 and calls it "our daily driver for agentic workflows." Carbon and Argon stay the deep-reasoning models. Pricing is $1.50 / $7.50 per million input/output tokens, flat across the full 1M-token context, so there's no surcharge for long prompts.

On the Artificial Analysis Intelligence Index, Google reports 51 at the default "high" thinking level against 50 for Gemini 4 Argon, and 54 at a new "xhigh" level. Against Gemini 3.8 Flash, which it replaces, Google says Hydrogen finishes tasks 2.5x faster and scores 57% on Terminal-Bench 4.1 versus 18.2%.

It's live in the Gemini API, AI Studio, Android Studio and Antigravity for paid AI Pro and Ultra subscribers. If you're still on 3.8 Flash, this is the migration to test first.

Meta Muse Ember: an even cheaper agent model

Why it matters: setups where an expensive model plans and a cheap one does the work only pay off when the cheap one is genuinely cheap. Ember is priced for that.

Meta released Muse Ember on November 9, its smallest and cheapest Muse model. Meta pitches it for high-volume agent work: tool calls, coding subtasks, routing, extraction and long-document parsing. Its line is "Muse Wonder plans, Muse Ember handles the volume." Vercel's AI Gateway lists it as meta/muse-ember at $0.10 / $0.50 per million tokens with a 1M-token context window. That's 15x cheaper than Hydrogen on input.

Meta says Ember is cheaper and smarter than Anthropic's Haiku and OpenAI's Luna. Those are Meta's own numbers, and TechCrunch, as quoted by Let's Data Science, notes they haven't been independently measured. At this price it's still worth a test on your simplest agent steps: extraction, classification, routing.

What to watch next: whether GitHub responds to the wfx campaign and whether more poisoned package releases turn up, plus independent benchmarks for Hydrogen and Ember against Claude Haiku 5.5.

Sources: The Hacker News — 4,837 hijacked GitHub accounts ran a CI secret stealer · Endor Labs — GitHub Actions supply chain attack targeting CI secrets at scale · CNBC — Anthropic releases Claude Fable 5.2 and Mythos 5.2 · Anthropic — Models overview · Google — Introducing Gemini 4 Hydrogen · SiliconANGLE — Google releases Gemini 4 Hydrogen · Let's Data Science — Meta releases Muse Ember · Vercel AI Gateway — Muse Ember

More from the blog