2026-10-22 digest
19 items on this page: 18 with their own page, 1 short.
A digest puts the short items from one day or one week on a single page. Opened alone, those briefs are often a headline and a sentence, which is a poor fit for a standalone article. Items that meet the importance threshold keep their own pages and are listed here by title.
Items with their own page
Google Cloud unveils Agent Compute Engine, the microVM platform behind Gemini agents: one hardware-isolated computer per agent, paused while it thinks, with no credentials inside; preview in the coming months
LiteLLM 1.107.0 on PyPI and Docker was poisoned with a credential stealer and RAT after attackers took GitHub tokens from an exposed internal LiteLLM deployment's logs, where LiteLLM says Claude Code had embedded them
Hugging Face open-sources Carbon-X, a 14B Apache-2.0 model you can chat with about DNA: averages 66.7% on five DNA benchmarks vs GPT-6.1 Sol's 67.5% at ~1/1,800 the cost, runs on one 80 GB GPU
Barclays to cut about 5,000 jobs (5% of staff) by year-end in an AI-linked efficiency push; 60% in support functions including developers, many in India
miHoYo and Tsinghua open GameScaling: eight Apache-2.0 game-playing models (0.8B–27B) that output keyboard-and-mouse actions; 27B tops Minecraft specialists and 9B beats frontier models in real-time play
iFlytek launches Spark X2.6: a 293B-A30B MoE with 1M context that it says was trained entirely on Chinese chips, priced at ¥0.9/¥3.6 per million tokens; only the 4B sibling has open weights
Straiker: Mexican crime group CyberCartel wired Claude, Grok and Kimi into a PentAGI-based agent swarm that hit 400+ Latin American banks, falling back to hardcoded scripts when models refused; Anthropic cut its access
Microsoft: China-state-aligned Storm-0940 used the ARTEX agentic AI framework against hundreds of organizations and in the ConnectWise supply-chain breach, hitting 2,250+ tenants at 2,900 sign-ins a minute
Walmart opens an MCP Agent Gateway for AI shopping agents, adds per-agent profiles with spending limits and terms barring card access, as Amazon keeps blocking Muse
Baseten's Base Labs open-sources decision models mammoth-1 (27B) and elephant-1 (9B) under MIT: mammoth-1 tops the Jev Decision Index at 71.24, elephant-1 ties Liquid's closed d1-ultra on Artificial Analysis
OpenAI's new usage policies put model routers and AI gateways behind prior approval and spell out a ban on using its outputs to build RL environments, rubrics and reward data for rival models; effective Nov 20
Meta creates non-voting Class C stock to fund the AI race as Q3 free cash flow turns negative on $34.9B capex; 2026 capex narrowed to $135–145B
Microsoft researchers resolve the 31-year-old KLS conjecture with human-directed AI agents; independent experts post a parallel proof calling their criterion the crucial step
Anthropic opens Claude Duet research preview: a standalone proactive-agent app on Opus 5.5 that briefs you daily, works in its own cloud computer and asks before it sends, pays or signs
Anthropic publishes its first Public Benefit Report, grading itself on five charter pillars and admitting gaps from test-environment escapes to no federal safety law; physicist Christopher Stubbs joins its Long-Term Benefit Trust
Claude Security moves to continuous scanning in Enterprise beta: auto re-scans as code changes, tracks fixes, editable threat models, plus API and MCP; Anthropic says users found 1M+ vulnerabilities since April
Anthropic ships Claude API version 2026-10-21, its first since 2023: temperature, top_p and top_k now return errors, new major SDKs switch to Bearer auth and Apache 2.0
Recursive releases Flint, an Apache 2.0 open-weight 40B-A2.5B MoE for local coding agents: self-reported 74.2% SWE-bench Verified and 150+ tokens/s on a 64GB RTX Spark, built with its Eureka Machine automated-research system
Short items
Cohere launches Rerank 5 Pro and Fast: tops BRIGHT at 82.2 (+5.8 over Voyage Rerank 3), adds Context Focus passage pruning that cut agent tokens 35% while raising accuracy, and probability-calibrated relevance scores
【Key points】
Cohere released Rerank 5 on Oct 21 in Pro and Fast tiers, generally available on the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker and North, plus VPC/on-prem. API price is $0.05 (Pro) and $0.02 (Fast) per 1M tokens.
1. Self-reported scores (RCP-nDCG@10): Pro ranks first on 39 of 46 composite datasets; 82.2 on reasoning-heavy BRIGHT, +5.8 over Voyage AI Rerank 3 and +12.4 over Rerank 4 Pro. Fast beats Voyage Rerank 3 Lite and the larger Qwen3-Reranker-8B on all four headline benchmarks.
2. Context Focus returns the most relevant passages inside each document in the same call. Cutting half the context kept ~96% answer quality; in a search-agent test the full stack used 35% fewer tokens than BM25 alone with accuracy up from 46% to 58%.
3. Calibrated scores read as probabilities (0.7 ≈ 70% useful); expected calibration error ~0.06 vs roughly 4x or more for rivals.
4. Fast handles ~187 1K-token docs/s on one A100, 2.9x Pro's throughput.
【Verifiable】
https://cohere.com/blog/rerank-5
https://x.com/cohere/status/2112923503100125326
【Our read】
1. Background: agents run dozens of searches per task, so retrieval quality and context length drive token bills; Cohere turns reranking into choosing which sentences reach the model.
2. Comparison: main rivals are Voyage AI (MongoDB) Rerank 3 and open Qwen3-Reranker. The differentiators are Context Focus (replacing a separate compression step) and calibrated thresholds that let agents tell when nothing relevant came back.
3. Impact: direct cost savings for enterprise search and support RAG, part of a wider shift, alongside OpenAI's longer prompt caching and Anthropic's cheaper cache reads, toward spending fewer tokens.
4. Caveats: Cohere's own RCP-nDCG metric isn't directly comparable to standard nDCG; the agent test used GLM-5.3 Flash; weights are closed.
Aggregated by AGI Pulse. Titles and links only; no full text is republished. Editorial policy