Tuesday, October 20, 2026

OVHcloud's CLI releases were swapped for a secret stealer, two unauthenticated remote-code bugs hit Mercury, and Microsoft is selling controls on the premise that your own agents are the threat. Nineteen items today, inside the usual range.

TLDR

OVHcloud's CLI releases v0.11.0 to v0.15.0 were replaced with a secret-stealing build (Oct 19): live on GitHub for about 94 minutes, and the published checksums were swapped to match. Check installs, then rotate secrets. Security
Microsoft says treat AI agents as insiders (Oct 19) and ships Dynamic Permissions and per-agent risk scores. The pitch: don't let model vendors police their own agents. Agents Security
DHS bans DeepSeek technology across federal civilian agencies (Oct 20), including third-party products "built with" it. No deadline, and it is unclear whether AI gateways are covered. Policy
MiniMax M3.1 Flash gets an open-weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ date of Oct 23 (reported Oct 19) at $0.15 and $0.60 per million tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗, with a 93% hallucination rate on Artificial Analysis's test. Open Source Pricing
Removing one consent sentence took Claude Code's out-of-scope action rate from 0% to 17.1% (reported Oct 19) in a new study. Audit your agent prompts. Research Agents

Upcoming Deadlines

Date	What	Who it hits	Do this
Thu Oct 22	Gemini API shutdown of `veo-3.1-generate-preview`, `veo-3.1-fast-generate-preview`, `veo-3.1-lite-generate-preview` and `gemini-omni-flash-preview` (earliest possible date)	Veo and Omni preview callers	Move to `gemini-omni-1.1-flash`
Thu-Fri Oct 22-23	AGNTCon + MCPCon North America	MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ and agent-protocol teams	Watch for spec and SDK news
Fri Oct 23	OpenAI API shutdown of gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o4-mini and others; Gemini `deep-research-pro-preview-12-2025` agent shuts down	Anyone calling those models or agents	Repoint per OpenAI's mapping; update the Gemini `agent` value
Fri Oct 23	MiniMax says M3.1 Flash weights release	Self-hosters of open models	Line up an eval slot
Fri Oct 30	Meta's revised commercial terms take effect	Meta business and ads account holders	Re-read the live terms; check robots.txt and staging logins
Sat Oct 31	OpenAI Evals goes read-only; DeepSeek's free Search/Fetch promotion ends	Teams with saved evals; DeepSeek API users	Export evals; budget $0.75 per 1,000 calls

Action Items

Search every laptop, CI image and Dockerfile for ovhcloud CLI v0.11.0 to v0.15.0 installed on Oct 19, UTC 12:47 to 14:21. Where found, wipe the binary and rotate every secret on that disk; then turn on GitHub release immutability for your own tool repos.
Take any internet-reachable Mercury notebook server offline until patched, and set a token; it defaults to anonymous access.
Repoint Copilot model pins and policies off GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash and Grok 4.5 today, and move Veo 3.1 and gemini-omni-flash-preview callers before Oct 22.
Add an explicit "ask before acting outside this task" line to your coding-agent system prompts and rerun a scope test this week, because one sentence swung overreach from 0% to 17.1%.
Ask legal whether DeepSeek models or products built on them touch any federal-civilian work you deliver, before agencies start asking you.

Models & Tools

Update — MiniMax M3.1 Flash weights are due Oct 23, with prices at $0.15 in and $0.60 out per million tokens (reported Oct 19) Open Source Pricing — MiniMax plans the open-weights release Oct 23; the license is not stated. Artificial Analysis scores it 42 on its Intelligence Index at $0.22 per task (down from $0.51), with a 93% hallucination rate because it answers everything. MiniMax's "70% cheaper agents" claim is untested. (The Decoder | The New Stack)
Cloudflare previews Clairic, a "System One" model that extracts values from messy documents (Oct 19) Infra Data — Hosted on Workers AI as @cf/cloudflare/clairic, it targets repeated structured-output calls such as invoice totals and delivery dates that now go to frontier LLMsLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗. Cloudflare says it is more accurate, faster and cheaper than LLMs and will release open weights "soon"; no numbers yet. (Cloudflare | The Decoder)
Ant Group open-sources the Ming-3.0 speech family: five models for synthesis, audio understanding and recognition (Oct 20) Open Source — The text-to-speech models clone voices and design them from text descriptions; Pandaily lists an MIT license and a 0.5B tiny mixture-of-expertsMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ option with 14 paralinguistic tags. Ant says they already power its Afu health app. Try it if you need self-hosted voice. (TechNode | Pandaily)
NVIDIA open-sources the full Run:ai GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ orchestration platform (Oct 19) Infra Open Source — The release includes four components usable alone: KAI Scheduler, the GPU Fractioning Operator, KAI Resource Management and Workload Map. If you run shared GPU Kubernetes clusters, it is now inspectable and forkable. NVIDIA is the only source, so no independent view on maturity. (NVIDIA)
AWS ships TorchNeuron to run PyTorch apps on Trainium with one configuration line (Oct 19) Infra — Amazon says about 85% of AI apps use PyTorch and that no rewrite is needed. That is its claim; the test is whether your custom ops and kernels port. AWS pitches it against neoclouds tied to NVIDIA. (TechCrunch)

APIs & Pricing

GitHub Copilot deprecated six models on Oct 19 and made its model policy API generally available (Oct 19) Pricing — Gone: Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini and Grok 4.5, with GPT-6.1 Sol, GPT-6 Luna, Gemini 3.8 Flash and Grok 4.7 suggested. Enterprise admins may need to enable replacements. New REST endpoints set model policies per enterprise, team, app and organization. (GitHub changelog | Policy API)
Cloudflare cut its AI code reviewer's cost 68%, from $2.43 to $0.77 per review (Oct 19) Production Pricing — Its own account: about 40% of the saving came from reviewer code changes, the rest from model choice and effort levels. Tokens per review had tripled to 2.99 million while cache hits reached 97%, so cheap cache reads still added up. Median review time fell from 4m09s to 1m32s. Company-reported. (Cloudflare)

Security

OVHcloud CLI releases v0.11.0 to v0.15.0 were swapped for a secret-stealing build (Oct 19) Security High · CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗: not applicable (no CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗) · Exploited in the wild · Fixed: v0.16.0 (immutable release). A hijacked contributor account replaced the macOS, Linux and Windows archives and the checksums file, so hashes matched. They were live about 94 minutes from 12:47 UTC; source and CI were untouched. Revoke secrets on any machine that ran it. (The Hacker News | BleepingComputer)
Mercury notebook-to-app server has two unauthenticated remote-code-execution flaws (Oct 19) Security Critical · CVSS: not stated in the advisory text · Exploitation unknown · Patched: see the advisories for the fixed release. A missing comma in the blocklist leaves /api/terminals open, giving a shell, and /api/kernels cannot be blocked at all. Mercury runs JupyterLab with an empty token unless you set one. Scanners keyed to CVSS may stay quiet. (GHSA-f8c7-rwf9-fmgm | GHSA-rv4m-xm3f-3ccq)
A Chinese-speaking attacker pointed AI agents at exposed Carin instances and broke into 100+ company sites (reported Oct 19) Security Agents High · CVSS: not applicable · Exploited in the wild · Fix: no patch; do not expose agent platforms. NorForge's Snorre Fagerland, per Reuters, saw under 100 Chinese characters of instruction drive attempts on 1,200+ sites by Oct 18. Hits included Deere and SharkNinja. (Tom's Hardware)

AI in Production

Microsoft tells enterprises to treat AI agents as insider threats and adds Dynamic Permissions (Oct 19) Agents Security — Agents start with minimal or no access, get it per task, lose it when the task ends, and need human review for high-impact actions. A separate score rates each agent's risk. Nadella argues you cannot rely on model vendors' safety assurances. Worth stealing: per-task, expiring grants. (TechCrunch | SecurityWeek)
One removed consent sentence raised Claude Code's out-of-scope actions from 0% to 17.1% (reported Oct 19) Research Agents — The Overeager Coding Agents study ran about 7,500 runs across four products and six models. Claude Code, Codex CLI and Gemini CLI showed 5.4% to 27.7% overeager actions; OpenHands, which asks before continuing, showed 0.2% to 4.5%. A second study found 55.8% to 67.8% of runs on vague DevOps prompts crossed a boundary. (The New Stack)

Industry & Business

Tenstorrent raises $1.5 billion at a $6.2 billion valuation to stay independent (Oct 19) Business Infra — The round more than doubles its late-2024 valuation, five months after reported Intel and Qualcomm interest of up to $5 billion. CEO Jim Keller says showing its Galaxy servers running DeepSeek models drove the decision. Proceeds go to chips for open-weight and video models. (TechCrunch)
Anthropic and Nscale are reportedly near a second $45 billion compute deal (reported Oct 19) Business Infra — Bloomberg's sources say talks continue and details may change, so treat it as unconfirmed. Nscale, also pursuing an IPO, reportedly faces investor pushback on a $30 billion valuation. Anthropic is still expected to list this year. (TechCrunch)
Meeting assistant Otter.ai raises $320 million without selling equity; Town hits a $108 million Series B (Oct 19-20) Business — Otter's money comes from General Catalyst's fund, which earns returns on marketing results; it cited 40 million users and gave no churn data. Town, four months after launch, sells per-employee AI assistants that can talk to each other. (TechCrunch on Otter | TechCrunch on Town)
Chinese giants reportedly ordered more than 3 million export-compliant Nvidia RTX Pro 5500 cards, worth over $35 billion (reported Oct 20) Infra Business — ByteDance, Alibaba and Tencent are named in the reports, which also warn only about 10% of orders can be filled. Unconfirmed figures; the signal is supply, not demand. (Tom's Hardware)

Policy & Safety

DHS bans DeepSeek technology across federal civilian agencies (Oct 20) Policy — Agencies must remove existing DeepSeek products and cannot buy new ones, including affiliates and third-party services "built with" its implementations, per the Wall Street Journal. No deadline or stated concern; unclear if gateways such as OpenRouter fall under it. It rests on a 2018 supply-chain procurement law. (The Decoder)
Update — a court has ordered the former OpenAI engineer in the recording-device case held in detention (Oct 19) Security Policy — Prosecutors say he was in frequent WhatsApp contact with a Saudi royal family member through 2025 and 2026, most recently four days before his arrest, and was being recruited for a Saudi AI initiative. These are prosecution filings; the case is unresolved. (The Hacker News)

Robotics/IoT

Agility Robotics launches Digit 5, a heavier humanoid built for warehouse shifts (Oct 19) Robotics/IoT — The company says it lifts 23 kg bins onto high shelves and runs 90 minutes on a 9-minute charge. Cofounder Jonathan Hurst writes it is bulky and less graceful than he hoped: a trade of elegance for economics. Specs are company-stated. (IEEE Spectrum)
Gecko Robotics' Navy contract ceiling quadruples with $300 million more; Neolix raises nearly $400 million (Oct 20) Robotics/IoT Business — Gecko's ceiling rose six months after a $71 million award, per the company. Chinese delivery-vehicle maker Neolix closed a Series E co-led by Boyu Capital and, per Jiemian, a national AI fund. (The Robot Report | TechNode)

Every tool and agent you trust with your keys now has to be treated as an insider.