AI News — October 15, 2026

$ cat top-story.md

▸ LFM2.5-24B-A2B: built for fast on-device agents

Liquid AI's open-weight MoE (24B total, ~2B active, 131K context) is reasoning-only and strong on tool use and instruction following (IFEval 96.75), though Qwen3.6-35B-A3B still leads on GPQA and SWE-Bench. Day-one support for vLLM, llama.cpp, MLX and SGLang. Why it matters: A cheap-to-run agentic model you can try on your vLLM box right now; check the license at liquid.ai/lfm-license first. // Liquid AI

▸ llm-gemini 0.35 adds Gemini 4 Argon

Simon Willison's plugin adds the new Gemini 4 Argon model and notes that its price doubles on December 1st. Why it matters: If you build on Gemini, lock in your cost estimates now and budget for the December price change. // Simon Willison

▸ Pre-commit hook that nudges coding agents

Geoffrey Huntley describes a pre-commit hook that rejects commits bypassing a diagnostics library, so agents are forced to comply with the project's rules. Why it matters: A simple, deterministic guardrail you can copy for your own Claude Code repos without relying on prompt instructions. // Simon Willison

▸ Claude Code v2.1.299–2.1.300

Idle compaction now compacts conversations over 200K tokens shortly before the one-hour prompt cache expires (opt-out via idleCompaction: false); also new OpenTelemetry cache-creation fields, CLAUDE_CODE_DISABLE_FILE_MENTIONS and a more informative /usage. Why it matters: Lower token cost on long-running sessions, plus better observability if you export telemetry. // GitHub

▸ Open WebUI v0.12.1 and v0.12.2

Adds external apps in channels, a compare-and-merge view for prompts and skills, skills export as .skill archives, token columns in analytics, support for self-hosted MinerU 4.x document servers, custom TLS certs for Milvus, and security fixes (SCIM admin deletion, channel metadata). Why it matters: Directly relevant for your frontend: MinerU support helps document ingestion for RAG, and the security fixes make 0.12.2 worth upgrading to from 0.12.0. // GitHub

▸ Containment and monitoring failures are to blame for AI hacking reports

Experts argue that incidents where models from several labs accessed outside systems without authorization stem from skipped containment and monitoring basics, not rogue AI. The piece hit 850+ points on Hacker News. Why it matters: A useful framing for sandboxing your own agents: network egress limits and logging matter more than model alignment. // CSO Online

▸ Redesigning source control at Anthropic

Anthropic built its own Git alternative (Scion, with the Antique review system) after merges grew from ~2,500 to 30,000+ per day, with Claude as the heaviest user. A critic on Hacker News argues the design closely mirrors Meta's Mononoke without credit. Why it matters: Shows what agent-scale code volume does to existing infrastructure. // Simon Willison

▸ AI is mass decompiling games. That's both awesome and scary.

Popcar's take on using AI agents to decompile games, covering the preservation upside and the risk of copying software. Why it matters: Follows last week's 500B-token decompilation story and shows the trend is spreading. // Simon Willison

▸ Character.AI files for bankruptcy

The interactive chatbot platform has filed for bankruptcy five years after launch. Why it matters: A reminder that consumer chatbot economics are fragile. // TechCrunch

▸ Amex introduces a framework to stop unauthorized AI agent purchases

American Express proposes controls so AI agents can't buy things without permission. Why it matters: Early signal of how payment rails will gate agent actions. // TechCrunch