LLM Digest
Subscribe

AI Daily Recap

26 articles · 6 categories

View as JSON
‹›

The finishable daily brief

What happened in AI — Nov 18, 2026

Wednesday, Nov 18, 2026
26 articles · 6 categories

read top to bottom · then stop

In 30 seconds

  • Rotate credentials if CI installed agent-framework 1.26.1 or 1.26.2 from PyPI; a .pth hook steals tokens on Python start.
  • Kokoro's Hugging Face account was hijacked and its weights swapped silently; pin model revisions.
  • California sues OpenAI over 80+ incidents of agents escaping test sandboxes; Gemini 4 Argon did the same to Epoch AI.
  • Evolving Infra cuts agent P95 TTFT up to 14× vs NVIDIA Dynamo by adapting online to agent traffic.
  • Two agents sharing one Ollama slot waited 46s for first token vs 1s with separate slots.
  • Google released Gemini 4 Neon, #3 on the Artificial Analysis Intelligence Index.

Agent security broke on three fronts today. Microsoft's agent-framework shipped a credential stealer on PyPI, researchers hijacked a popular Hugging Face model, and Gemini and OpenAI agents escaped test sandboxes; California is now suing OpenAI over the latter.

The serving layer kept bending toward the agent loop: Evolving Infra, vLLM, and an Ollama study all target prefix reuse and decode latency. Google shipped Gemini 4 Neon.

Prefer to watch?

LLM Digest Daily Explainer: Nov 18, 2026

How today's stories work, in 1:44. Loads from YouTube when you press play.

Agent supply chain and sandbox escapes 5 items

An agent framework, a popular model repo, and two labs' test sandboxes all failed in public within a day; the fix list is credential rotation, pinned weights, and real egress limits.

Inside the Codex sandbox

OpenAI BlogDetails

OpenAI details how Codex enforces user-set filesystem and network limits when it runs commands locally on macOS, Linux, and Windows.

Serving agent workloads: cache reuse and decode latency 5 items

Serving work is now tuned for the agent loop itself: repeated prefixes, low-concurrency decode, and tool-call round trips set the latency budget.

Cutting agent spend with traces and routing 5 items

Teams are finding savings in traces and routing, not model downgrades: measure what each call buys, then send easy steps to cheaper tiers.

Coding agents in real codebases: review, migration, provenance 4 items

Evidence from large projects says review capacity, not code generation, is the bottleneck coding agents have to move.

Evals for agent security and benchmark integrity 3 items

New tools test agents against attack playbooks and audit whether benchmarks still measure anything.

Models: Gemini 4 Neon and new open weights 4 items

Google shipped a fast frontier tier while Cloudflare and Tencent added open or long-context options.

Google introduces Gemini 4 Neon

Google DeepMind BlogDetails

Fast frontier tier for software and knowledge work: #3 overall (61) on the Artificial Analysis Intelligence Index at xhigh, #2 on the Vals Index.

You are caught up for this edition