OpenCode disclosed a breach of its older Console system affecting 1.4% of users

OpenCode disclosed a breach of its older Console system affecting 1.4% of users, calling it a 'large-scale distillation attack.' Some individual accounts saw unauthorized AI-provider charges exceeding $1,000. OpenCode's own AI models refused to analyze the attack code on safety grounds during the 48-hour incident, echoing a July 2026 Hugging Face guardrail-asymmetry problem.

Nscale and Anthropic are reportedly in talks for a second ~$45B compute deal for a new West Virginia data center, barely a month after their first $45B/six-year deal (August 2026) — a potential ~$90B combined commitment, pending signature.

Microsoft expanded Guardian Agents under Project Perception, extending its 'AI agents are insider threats' framing to 60 monitored local-agent types and separate Entra Agent identities — while remaining an investor in and cloud provider to the same vendors (OpenAI, Anthropic) it is selling agent-security products against.

A Microsoft-authored paper finds an agent harness's context-management choices, not the model, can decide task success: dropping an irretrievable fact costs a 92.5-percentage-point penalty versus a re-fetchable one, and summarizing old turns recovers 86% of a 63-point ARC-AGI-3 harness-design gap.

An independent paper proposes honeypots — decoy harmful-action opportunities planted in live production — as an AI-agent insider-threat control that deters misbehavior through uncertainty alone, separate from detection, and stays useful even as agent reasoning grows less interpretable.

akitaonrails/ai-memory climbed to #3 on GitHub Trending and got its first deep-read: a Rust MCP/HTTP server giving 20+ agent CLIs (Claude Code, Codex, Cursor, Gemini CLI) one shared, git-versioned memory store via a typed handoff protocol, MIT-licensed, with a 'zero LLM calls' default capture path.

Claude by Anthropic is absent from the captured Google Play top 30 for a fourth consecutive day, extending the reversal first logged 2026-10-17 toward the page's own week-long bar for confirming a genuine, sustained drop.

Fish Audio's drama-3 (35B-parameter MoE, 3B active, 16-speaker voice cloning, sub-100ms serving latency) is the strongest synthesis-only TTS artifact logged in weeks, but its research-only license and lack of full-duplex speech-to-speech leave a standing voice-tech suspension test unmet for a tenth consecutive rotation.