Tuesday, 20 October 2026

Notion has rebuilt its AI agents and moved Notion AI to a new credit system [1]. Business and Enterprise seats now include 1,500 Standard credits a month, for work where Notion's Auto router picks the model, and 500 Select credits a month, for work where you pick the model or run your own Custom Agents [1][2]. The system took effect for new workspaces on 19 October; existing workspaces move over on 2 November [2][3].

The two allowances are separate for each member, cannot be pooled or transferred, and do not roll over [2]. When one runs out, the features it covers pause until the monthly reset, unless an admin has turned on workspace Notion credits, which is off by default [2]. Those workspace credits also pay for high-consumption models such as Fable and Astra, and for Custom Agents the workspace funds [2]. AI Meeting Notes will start drawing on Standard credits on 1 December, at about 10 credits for a one-hour meeting, Notion estimates [2].

The company says the upgraded Auto router delivered results similar to Opus 5.5 and Astra across 20+ common tasks at about 90% lower average cost [1]. Its help page estimates a simple task at 5 credits on Auto against 50 with Opus 5.5 [2]. Agents can now work on a single task for more than 24 hours, up from five minutes, and get their own computer and browser plus a "Review for me" mode that uses a separate model to check actions [1][3]. Members can connect a ChatGPT Plus, Pro or Go plan to their personal Notion Agent, and Enterprise customers can test bring-your-own-key in private alpha [1][3].

Notion says the vast majority of users are unlikely to run out of monthly credits [1]. How many credits a heavy user actually burns in a month is not public.

The US Department of Homeland Security has ordered federal civilian agencies to stop buying DeepSeek products and to remove them from their IT systems [1][2]. The notice, published in the Federal Register on 20 October, covers products and services from Hangzhou DeepSeek Artificial Intelligence Co. and its affiliates, and also "third-party products and services developed using DeepSeek implementations" [1]. That second clause is the wide part: the order reaches beyond DeepSeek's own app to products built on its technology [1][2].

The legal basis is the Federal Acquisition Supply Chain Security Act of 2018 [1]. According to [3], the notice publicizes Exclusion and Removal Order 2026-01, signed by Homeland Security Secretary Markwayne Mullin and originally released on 24 July. The exclusion took effect on issuance, and the 60-day window for agencies to submit written exception requests, which must name the affected products and give a compelling mission or national-interest reason, runs from that issuance rather than from the Federal Register publication, according to [3]. The secretary must review the order within one year and annually after that, according to [3]. Questions go to the FASCSA Order Implementation Team in CISA's Cybersecurity Division [1].

DHS raised concerns that sensitive information from US government users could be accessible to Chinese authorities under Chinese law, according to [2]. Cyberpress lists the threat vectors DHS identified as sensitive information exposure, user tracking and application security weaknesses [3]. China's foreign ministry said Beijing opposes "overstretching the concept of national security and politicizing technology and trade issues" [2]. A law passed in December 2025 already targeted DeepSeek's use in intelligence national security systems; this order extends the restrictions to civilian agencies, covering both new purchases and existing systems, according to [2].

The order binds federal agencies, not private companies. But any vendor selling software to those agencies now has to know whether a feature in its product runs on a DeepSeek model.

Otter.ai, the AI meeting notetaker, has closed $320 million in growth financing from General Catalyst's Customer Value Fund [1]. The fund provides capital to startups without diluting equity and gets its return on the results of marketing spend, according to [2]. Otter says the money will go to enterprise growth, go-to-market strategy, international expansion and product work [1][2].

The company says it has more than 40 million users, including teams at over 86% of the Fortune 500, and has transcribed more than 1 billion meetings [1]. TechCrunch puts the user count at 25 million a year earlier, and reports that Otter crossed $100 million in annual recurring revenue last year [2]. Before this deal, Otter had raised $50 million in total, with its last round in 2021, according to PitchBook data cited by [2]. Otter did not give updated revenue figures, monthly active users or churn [2].

CEO Sam Liang says the financing will help Otter "build enterprise agentic workflows" [1]. TechCrunch reports the company is testing a pre-meeting brief built from earlier conversations and a Live Assist feature that lets users ask the assistant questions during a meeting [2]. The market is crowded: TechCrunch lists Fireflies.ai, Granola and Read AI, newer entrants such as Calendly and Wispr, and Superhuman, which bought Fathom [2].

For a small team that already records calls with Otter, nothing changes today. The direction is toward enterprise features and agent workflows built on meeting transcripts. The terms of the financing, including what Otter pays back and when, are not public.

Mistral has released Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice agent SDK that puts listening, reasoning, tool calling and speaking into one real-time streaming pipeline [1]. It is available through Mistral's EU-hosted API or for self-deployment on your own infrastructure [1]. According to [2], the SDK is built on LiveKit.

The prices are published [1]. Voxtral Mini Transcribe 3 costs $0.0035 a minute in batch, Voxtral Small Transcribe 3 costs $0.0040 a minute, and Voxtral Mini Transcribe Realtime 3 costs $0.008 a minute [1]. Voxtral Speech 3, the text-to-speech model, costs $13 per million characters for batch and realtime use [1]. The company says that is less than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash [1].

Transcription now covers 50 languages, including the 24 official EU languages, with speaker diarization and context biasing, which lets you feed in key terms to improve accuracy without fine-tuning [1]. The Mini model has 4B parameters [1]. On Mistral's own chart, Voxtral Small Transcribe 3 scores a 4.0% word error rate on the FLEURS benchmark across 45 languages, ahead of Whisper large-v3, Gemini 3.5 Transcribe and Scribe v2 [1][2]. The realtime model reports latency below 500 ms [1][2]. Speech output covers 35+ languages and 50+ dialects, and a Pro Voice option reproduces a speaker's voice across languages from 30 minutes of audio [1].

The benchmarks are Mistral's own. If you run a phone line or a support desk, test it on your own calls, accents and product names before you switch.

OpenCode, an open-source AI coding agent, says attackers broke into an older system and exposed provider API keys that some users had stored in its online platform, OpenCode Console [1][2]. The company said the breach was "related to a large-scale distillation attack that impacted 1.4% of OpenCode Console users," and that it emailed all of them [1]. According to [2], OpenCode confirmed the breach on 17 October, and its email said a "legacy system that we were in the process of decommissioning" had been accessed.

The bills arrived fast. Developers posted spending alerts and unauthorized charges, with some reporting hundreds of dollars on individual accounts and a few above $1,000, according to [1]. Anthropic and Google blocked keys they believed were exposed, and OpenAI said it would refund its API customers for credits used in the attack [1]. OpenCode says refunds for unauthorized charges are underway [1]. Co-founder Dax Raad says the open-source app and command-line tool were not affected, because "those keys never come to us" [1].

Defending was hard too. Raad said attackers had tried to break in nonstop for 48 hours, and AI models the team used for analysis started refusing once the malicious code tripped their safety filters [1]. OpenCode later got access to OpenAI's Daybreak Blue program for vetted defenders [1]. How attackers got in and how the keys were stored has not been disclosed [1].

The checklist from supply-chain security firm Safeguard, as reported by Infosecurity Magazine, is short: list every key ever stored in a Console workspace, check usage and billing from 16 October, create replacement keys, then revoke the old ones and confirm they no longer work [2].

Microsoft has extended Project Perception, its security product, to treat AI agents as potential insider threats [1]. The company announced new Playbooks and a team of Guardian Agents that carry out protective actions at machine speed, inside boundaries people set [1]. Its security business chief, Hayete Gallot, put the premise bluntly: "frontier models, regardless of their origin, cannot be trusted," because they are unpredictable rather than malicious, according to [2].

The design follows a line from CEO Satya Nadella: separate the supply of intelligence from the authority over it [1][2]. Microsoft says the controls that decide what an agent can access, what it can do and whether it keeps running must sit outside the model [1]. Its example is an expense-approval agent that should not inherit every permission of the employee who sponsors it. If that agent meets a prompt injection, the system can keep safe reads while blocking writes, require human approval for a consequential step, or isolate and stop it [1].

Discovery comes first. Microsoft is adding 35 more types of local agents to its detection, for a total of 60, including Claude Code, OpenAI Codex and GitHub Copilot [1]. TechCrunch notes the position is partly self-serving: Microsoft has invested in both OpenAI and Anthropic, and wants enterprises to buy agent security from the cloud provider rather than from the labs [2]. Pricing for Guardian Agents is not public.

The idea scales down. A small team without Microsoft's security stack can still give each agent its own narrow permissions, log what it does, and require approval before any write that matters.