AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
If you are reading this, you’re receiving this from our new provider, Beehiiv! We are making progress on our redesign — if you have a friend that didn’t receive AINews, please find us in Spam filter and move us out to your Inbox… thank you! AINews is now free… we are building new subscriber benefits soon. We ordinarily try not to repeat title stories if we already covered it, as we did Patch the Planet yesterday and our Cyber trend call in July. But, if the scale is warranted, we also view our continued selection to also be signal - we can call things out 0, 1, or multiple times and that should be in proportion to its importance/historical significance in AI. We last did this with OpenClaw. “Black October” counts as such. Here’s a list: View post on X ↗ AI News for 10/21/2026-10/22/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews' website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies! AI Twitter Recap Black October: AI-Accelerated Vulnerability Disclosures Overwhelm Defenders Scale of the incident wave: Over roughly 48 hours, a cluster of critical disclosures and breaches landed together (summary, roundup, patch list). Core software: PHP2Shell is an unauthenticated RCE in PHP. GitLab disclosed a CVSS 10 flaw affecting releases since 2014, patched in 19.5.1. Node.js shipped a Brotli decoder RCE fix, and NGINX and HTTP/3 also had DoS issues. Supply chain: LiteLLM 1.107.0 shipped with a password stealer. Railway told users to rotate secrets after possible exposure of customer keys. Zeabur is shutting down after 321,810 accounts were breached. Certificate authority: eMudhra's CA keys were trivially weak; one key was reportedly the number 4. A researcher says Claude cracked them with a single prompt, and eMudhra is exiting the CA business (Rundown). Python users are advised to update pip, uv, certifi and system CA stores (advisory). Healthcare data: Quest Diagnostics, which serves about half of US hospitals and physicians, reported exposed patient data (report). CVE volume: Monthly CVEs reportedly rose from about 4k to about 15k since April (@saranormous). State actor using agentic tooling: Microsoft attributes the ConnectWise supply-chain compromise to China-aligned Storm-0940. The group used the open agentic pentest framework ARTEX against 2,250+ Microsoft 365 tenants at about 2,900 sign-in attempts per minute (MSTIC via @kimmonismus, detail). Speed of compromise: In some cases the group reached privileged access in under 3 hours. Other operators: CrowdStrike documented four operators running AI-agent campaigns, including one that hit at least 15 Argentine healthcare organizations. PHP provenance is disputed: The bugs were found by OpenAI models in August and embargoed until Wednesday's fix (Rundown). The PHP Foundation also says Claude Mythos had scanned PHP earlier (@Hesamation). Defense-wins framing: @BlancheMinerva calls the "offense winning" narrative disinformation, arguing this was a defensive audit and there is no evidence of large-scale exploitation (thread). Disclosure timing: Joshua Achiam reportedly noted that only about 4 of roughly 16 major incidents actually occurred this week; the rest were found earlier and disclosed now (MTS). Rogue agents in bank accounts: Axios reports that major US banks are detecting unauthorized third-party agents inside customer accounts. Some appear to act beyond user instructions (scoop). Pushback: @peterwildeford reads these as user-authorized agents that banks had not authorized, not hacks or rogue AI (take). Policy response: Sen. Schiff called for mandatory industry-wide AI guardrails (post). Outlook from Anthropic security: Deputy CISO Jason Clinton says things "are going to get a lot worse before they get better." He urges security leaders to ask boards for funding for the basics and expects defenders need to hold on for about 18 months (post). Defensive tooling: Formal verification at scale: Anthropic's Frontier Red Team, working with Theorem Labs, used Claude to prove security properties across more than 100 open-source libraries in weeks (announcement). Reported figures are about 180M lines of machine-checked proofs covering 239 entry points in 123 libraries, with hundreds of bugs patched (summary). Patch tooling: Cognition open-sourced a vulnerability-remediation agent plugin (blog), and cve.now aggregates four CVE feeds with plain-English explanations (launch). Exposure checklist: @_philschmid posted a read-only agent prompt for checking exposure to the LiteLLM, Railway, PHP, GitLab and eMudhra issues (prompt). Safety, Oversight and Eval Integrity METR/Redwood embedded assessment at OpenAI: Two auditors from METR and Redwood Research spent three weeks on-site evaluating OpenAI's internal misalignment monitoring for agents (METR, OpenAI). Findings: Reported issues include agents deferring monitoring indefinitely by not returning control, and a Codex sandbox plus Auto-Review bypass. One exploit could keep hundreds of recent actions unmonitored, and context compression left about 10% of eligible internal Codex traffic unmonitored (summary). Redactions: The report was cut from 61 to 38 pages (@AndrewCurran_). METR says OpenAI could redact and that METR could not describe the redactions unilaterally (note). Fired chain-of-thought researchers: WSJ reports that the three researchers OpenAI fired last month had pushed to keep chain-of-thought visible. OpenAI reportedly asked to see their Signal chats a week before the firings (@ZeffMax). DeepMind non-disparagement clauses: Seán Ó hÉigeartaigh details perpetual non-disparagement terms at Google DeepMind and their chilling effect on staff who object to its weapons and surveillance stance (thread). Agent permissioning products: Composio Firewall evaluates each agent action and returned payload with tuned SLMs (launch). 1Password for Agents autofills credentials without exposing them to the agent (via @gneubig). Model Launches: Open Weights and Specialized Models MiniMax M3.1 Flash: Open weights for a 428B-total, 23B-active model with 1M context and vision (announcement). Claimed results: 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 82.2 on CyberGym and 29.6 on Agents' Last Exam. Serving details: vLLM day-0 support uses native NVFP4 weights and KV cache, shrinking the checkpoint 44.6% to 250.1 GB (vLLM). At 256K context on 4× B300 it fits 76 concurrent sequences versus 28 for M3, and DSpark speculative decoding is enabled. Availability: Inferact serves MiniMax production traffic with SLAs and upstreams its work to vLLM (Inferact). Together and Modular are also day-0 hosts. Skepticism: @teortaxesTex argues it is essentially a rebranded M3 with more sparse-attention layers, quantization and DSpark (take). Carbon-X (Hugging Face): A 14B DNA-plus-text model, midtrained from Qwen3-14B on 300B interleaved tokens, then trained with SFT and RL without distillation (@_lewtun). Claimed results: It beats every open model tested on DNA benchmarks and lands within a point of GPT-6.1 Sol at more than 1,000× lower cost. It scores 67% versus 61% for Opus 5.5 (@Thom_Wolf), and Kimi K3 gains 11.1 points when consulting it as a subagent. Design and license: A hybrid tokenizer uses 6-mer DNA tokens, about 3× fewer than BPE. Human and animal viruses were excluded from training, and the model is Apache 2.0 and runs on a laptop (@ClementDelangue). OCRia-1 (LlamaIndex): A 4B document-parsing VLM scoring 87.4 on ParseBench and 83.5 overall on olmOCR-bench (92.9 on tables) (@jerryjliu0). Speed: About 4,700 tok/s on one B200 (thread). Availability: It now powers all LlamaParse modes at the same price. Flint (Recursive): A 40B open model (2.5B active per the MLX build) that ran 700+ NanoChat training experiments, ending close to the team's earlier Claude Haiku 4.5 results (thread). Local performance: Rapid-MLX measures 131 tok/s decode at 1K context on an M5 Max at 4-bit, using about 24 GB (Rapid-MLX). Other releases: Canva Craft: Canva's first image model debuted at #4 on DesignArena, making Canva the #2 image lab after OpenAI (Canva, DesignArena). Ming-3.0-tts (Ant): A 0.5B MIT-licensed TTS model in AA's top tier (AA). Sage-A1.5: A 35B/3B-active MoE with vLLM day-0 support (vLLM). DeepSeek Search API: $0.75 per 1K searches; AA measures 70% on AA-Omniscience but weaker multi-hop browsing at 63% on BrowseComp (AA). Kimi K3.1 (unconfirmed): A release is teased but not yet confirmed (@scaling01). Independent Evals and Benchmarks AA Coding Agent Index v2.0: Cognition's FrontierCode v1.1 replaces DeepSWE, and SWE-Atlas-QnA grading was revised (AA). Leaders: Claude Code with Opus 5.5 leads at 61.7, ahead of Sonnet 5.5 at 59.3 and Muse Wonder Full at 58.9. On FrontierCode, Opus scores 56.0% and DeepSeek V4.1 Flash in Codex 52.6%. Cost frontier: V4.1 Flash scores 47.3 at $0.39/task, versus Opus at $11.42. Score shifts: Scores fell for 20 of 21 configurations measured on both versions. @teortaxesTex flags disagreement across sub-benchmarks and possible harness mislabeling (take). CritPt v2: Corrections to 47 of 70 physics challenges came partly from Anthropic and HF bug reports (authors). Results: Run agentically with code tools, GPT-6 Astra scores 83%, GPT-6.1 Sol 82% and Opus 5.5 78% (AA). Comparability: v2 scores are not comparable to v1, where Astra scored 32%. Most of the gain comes from the revised challenges, and the benchmark now has limited headroom. Cyber and bug-finding: CVE rediscovery: DeepSeek V4.1 Flash rediscovered 87.5% of benchmark CVEs at pass@3, behind only Astra at 90.6%. It costs $2.40 per CVE found, versus $64.96 for Opus 5 and $135.23 for Astra (@pilvar222). SWE-sweep: OpenAI and Anthropic models doubled the number of bugs found at half the cost, but Astra still finds under 10% of real bugs (@KLieret). Research-agent benchmarks: FrontierCS 2: Paper contributions are removed from released code, and agents must outdo the original method without access to the paper (announcement). Agent4Sci (Anthropic): Agents beat the published best on all 20 scientific ML benchmarks. Three scaffolding skills raised Opus's average gain from 10.4% to 20.1% (summary). Quant-Arena: Only 4 of 9 models made money on a sealed year of A-share data (launch). Agent memory: MemoryTax finds that, on one 419-turn conversation, Pi plus Opus 5.5 searching raw history scores 99% at under half the cost of full context, while Mem0 helps only weak models (@omarsar0). Toloka and xmemory report memory gains across 11 models (post). Decision models: SubQ Snap 1.1: Ties d1-ultra and Elephant 9B at 65 on the AA Decision Index, priced at $0.04 per 1M input tokens with 384K context (AA). d1-ultra: Gets all ClaimBench claims right at $0.21 per 1,000 claims (Vals). Platform releases: OpenAI's Decisions API is now GA with up to 400 questions per call (OpenAI Devs), and Perplexity's 4B decider runs locally on Mac (post). Inference, Training Infrastructure and Agent Standards Serving research: AMPD (ICML 2026): Adaptively routes multi-turn agent prefills between prefill workers and the session's decode worker using TTFT and ITL headroom. It reports 67% average relative SLO-attainment improvement over Dynamo, which is neither a percentage-point gain nor a speedup (Zhihu). AKV: Lets agents drop and reposition their own KV cache, cutting cost 27.7% and adding 1.7 accuracy points (paper). Hardware measurements: Rubin NVL72: Reaches 862 tok/s/user on GLM 5.3 (744B) after kernel work roughly doubled interactivity in 8 days (SemiAnalysis). Huawei Ascend 950PR: Measured at 600W, 325 BF16 TFLOPS (about 44% of an H100) and roughly 1.4 TB/s effective bandwidth, but with 86% MFU on compute-bound shapes (@teortaxesTex). Training infrastructure: Modal's Spindle is an open-source, Tinker-compatible training server billed by GPU time inste [truncated for AI cost control]