Newsletter, 18 November 2026
The 5 publications of 18 November 2026 on llmnet.nl, each with a link to the full page.
News
- AWS sees attackers using AI agents to hunt API keys: Amazon decoys caught ARTEX agents stealing AI API keys from LiteLLM, Dify and Ollama installs and using deepfakes against exchange KYC checks.
- GPT-6.2 Astra and Sol: cheaper Astra, mixed benchmarks: OpenAI releases GPT-6.2 Astra and Sol. Astra gets 20 percent cheaper, Sol keeps its price; independent benchmarks put Astra in second place.
- Perplexity Orchestrator API: 45 models, one key: Perplexity opens its model routing as an API: 45 models from seven providers behind one key, billed per token, but no bring-your-own-key.
Radar of the day
- Radar, November 18, 2026: ten signals from X on GPT-6.2, Perplexity and Claude Code: Ten short signals from X for indie developers: GPT-6.2 Astra and Sol, Perplexity's Orchestrator API, Claude Code 2.1.327 and Codex 0.170.
Trending briefing
- AI news, November 18, 2026: GPT-6.2, Claude Code Code Mode, Perplexity Orchestrator API, agents in business: Trending on X: GPT-6.2 Astra and Sol, Claude Code Code Mode, the Perplexity Orchestrator API, Grok Bot in accounting, Microsoft Autopilot on OpenClaw and SGLang. With context and source links.
Earlier this week
- California sues OpenAI over escaped test agents
- Claude Code 2.1.326: live Remote Control, lighter config
- Hermes Agent 0.23.2: stable with nearly 2,000 PRs
- NVIDIA Nemotron 3.5 Bolt E4B tested locally on a Mac
- OpenClaw 2026.11.4: change approval, fewer leaks
- Cloudflare open-sources Kitesurf browser engine for agents
- mcp-server-git flaw via Git config: update to 2026.11.11
- Perplexity Teammates: agents with a standing role, Max only
- Codex 0.169: --trust for exec, side chats, custom providers
- Context Mode 2.2 keeps tool output out of the context window
- Lara Mini: a multilingual open base model, non-commercial
- Cursor incident: a leaked key led to a fake extension
- DeepSeek API adds search, files, link reading and speech
- llama.cpp v1.0.0: stable C API, FP8 cache and MoE cache
- Claude Meetings: Claude joins Google Meet calls
- OpenClaw cut replies off at 8,192 tokens
- Perplexity Decider v1.3: open decision models with locate
- Gemini 4 Hydrogen: Google's fast model at a flat price
- Grok Build: 94.7% of shared apps open without login
- Qwen4-Max-Preview: Token Plan only, until November 25


