AI News Roundup – October 19, 2026

TL;DR

• The US Department of Homeland Security has ordered DeepSeek products removed from federal civilian agency IT systems and barred from future purchases. The order also covers third-party products built on DeepSeek.
• Reuters: a Chinese-speaking hacker used AI agents to break into websites of more than 100 companies, including Deere, SharkNinja and Cummins, for roughly $50 in tokens per successful hack.
• Microsoft says AI agents are now “agentic insiders” and is extending Project Perception with Guardian Agents, Dynamic Permissions and agent risk scoring.
• MiniMax launched M3.1 Flash at $0.15 / $0.60 per million input/output tokens, claiming agentic-task costs 70% lower than its predecessor.
• Inception released Mercury 3, a diffusion LLM running at 1,068 tokens per second with a 1M-token context.
• Tenstorrent raised $1.5 billion at a $6.2 billion valuation to scale chips for open-weight models.
• doubleAI unveiled Maxwell, an AI that speeds up code while it runs, alongside a $200 million Series A.

---

Why is the US removing DeepSeek from federal systems?

Short answer: A Department of Homeland Security notice, filed on October 19 for publication in the Federal Register on October 20, 2026, puts on record an order from Homeland Security Secretary Markwayne Mullin that does two things. It excludes IT products and services from Hangzhou DeepSeek Artificial Intelligence Co. and its affiliates from federal civilian agency procurements, and it orders the removal of all DeepSeek and DeepSeek-affiliate products, plus third-party products and services developed using DeepSeek implementations, from civilian agency IT.

The order uses the Federal Acquisition Supply Chain Security Act of 2018 and follows a review by the Federal Acquisition Security Council. According to CyberScoop, the signed order cites DeepSeek data stored on servers in China, the collection of personal information, and the risk that the “prompt and response” process “may result in the transfer of sensitive data.” CyberScoop also reports that the signed order predates the public notice by months. Agencies can request an exception 60 days in advance with a “compelling justification,” and they are encouraged to monitor traffic and tighten logging and data-exfiltration controls in the meantime.

The order applies to federal civilian executive-branch agencies. It is not a ban on businesses or consumers using DeepSeek.

Why it matters: The order is government-wide, names a single AI lab, and reaches products built on DeepSeek, not just DeepSeek’s own app and API. If you sell software to US civilian agencies, you’ll need to know exactly which models sit inside your product. See DeepSeek and compare alternatives on our Models page.

Sources: Federal Register notice (PDF) · Federal Register public inspection · CyberScoop · WSJ · Bloomberg

---

How did one hacker use AI agents to breach 100+ companies?

Short answer: In a Reuters exclusive published October 19, NorForge principal threat researcher Snorre Fagerland described watching hackers through about two dozen exposed instances of Cairn, an open-source AI problem-solving platform, that were being used maliciously from around October 2. The most prolific, a Chinese-speaking hacker, broke into websites belonging to more than 100 companies, including Deere, SharkNinja and Cummins.

Starting September 25, the hacker pointed AI agents at US, Canadian, Australian and Japanese e-commerce sites, using instructions of fewer than 100 Chinese characters. The agents attempted more than 1,200 sites, and successful intrusions typically took about 23 hours of probing. Fagerland estimated the campaign, which appeared to use a mix of DeepSeek and Claude Code, used roughly $5,000 of tokens, or about $50 per successful hack.

Deere told Reuters it had “no indication that sensitive information was accessed.” SharkNinja said a standalone product-support site was hit and the threat was contained. Cummins said it was investigating a site run by a third-party vendor. Reuters could not determine how much data was taken. Cairn’s author says the platform was not designed for hacking.

Why it matters: This is real-world evidence that off-the-shelf AI agents can run a whole intrusion with almost no human effort, and do it cheaply. Public-facing control panels and vendor-run sites are the soft spots, so patch and monitor them. The same agent tools developers use every day, such as Claude Code, can be turned to this kind of work, so assume attackers have them too.

Sources: Reuters via Honolulu Star-Advertiser · Reuters via Yahoo News · The Verge

---

What are “agentic insiders,” and how does Microsoft want to secure them?

Short answer: In an October 19 post on the Official Microsoft Blog, Microsoft Security EVP Hayete Gallot argued that AI agents deployed inside companies are now “agentic insiders”: they are frequently overprivileged and can act in critical processes. She wrote that “frontier models, regardless of their origin, cannot be trusted,” because training for results leads to reward hacking. Her answer is Guardian Agents that “watch and moderate every action of every agent, before it happens, 24/7.”

Microsoft is extending Project Perception, its agentic security system announced in July, to do exactly that. A companion Microsoft Security Blog post lists the details:

• Agent discovery now covers 60 types of local agents (35 newly added), including Claude Code, OpenAI Codex and GitHub Copilot.
• Microsoft Entra Agent ID is being extended so every agent gets its own identity, separate from the person who uses it.
• Dynamic Permissions start agents with no or least privilege, scoped to the task and time-bound.
• Dynamic Agent Risk and Dynamic Conditional Access can respond step by step, for example keeping read access while blocking writes, or requiring human approval for a high-impact action.

Microsoft also wants security tools to see models’ chain-of-thought tokens, while keeping enforcement outside the model. Satya Nadella is quoted: “The most trustworthy agentic system will not be the one with the model we trust most. It will be the one designed that enables us to trust the model least.”

Why it matters: If your team runs coding or business agents, treat them like employees with access badges: separate identities, the least privilege that works, and a log of every action. See Microsoft Copilot, Claude Code, OpenAI Codex and GitHub Copilot.

Sources: Official Microsoft Blog: Securing the agentic insider · Microsoft Security Blog

---

How cheap is MiniMax M3.1 Flash?

Short answer: MiniMax officially rolled out M3.1 Flash on October 19 with flat pricing of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. MiniMax says the anonymous “Space Bunny” model, which processed more than 200 trillion tokens, was an early version.

MiniMax says overall usage cost on agentic tasks is 70% lower than its predecessor and online inference is nearly 3x faster, with output speeds of 200+ tokens per second (median 170+). Its self-reported scores include 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 82% on CyberGym. M3.1 Flash is available through the MiniMax API, M Plan and the MiniMax Code desktop app. Model weights will be released later, and a larger Pro model is also coming later.

Why it matters: Cheap, fast agent models keep driving down the cost of running coding and computer-use agents. The benchmarks are MiniMax’s own, so test it on your own workloads before switching. Compare current API prices on our Models page.

Sources: MiniMax: M3.1 Flash · MiniMax on X · Reuters via Investing.com

---

What is Mercury 3, and why is it so fast?

Short answer: Inception launched Mercury 3 on October 19 and calls it the most capable diffusion-based LLM released to date. Diffusion models generate text in parallel rather than one token at a time. Inception says Mercury 3 runs at 1,068 tokens per second on NVIDIA GPUs and matches or beats GPT-6 Luna and Claude Haiku 5.5 on its benchmarks while running 4–8x faster.

The context window grows to 1M tokens, up from 260K for Mercury 2.5, and the model is more than 40x larger than Mercury 1. List pricing is $0.20 / $0.70 per million input/output tokens, 50% off at launch ($0.10 / $0.35), with cached tokens 90% off the input price. It’s available on Inception’s API (with 100 million free tokens to try), OpenRouter, Baseten and Vercel, and it’s OpenAI API compatible.

Why it matters: For voice agents, support bots and browser agents, waiting on the model is the main source of lag. If diffusion models keep quality while staying this fast, real-time AI gets noticeably snappier. See the Models page for how today’s fast models compare.

Sources: Inception: Introducing Mercury 3 · Vercel changelog

---

Why did Tenstorrent raise $1.5 billion?

Short answer: AI chip company Tenstorrent closed a $1.5 billion Series E at a $6.2 billion post-money valuation, led by Trace Capital and Innovation X. Other investors include ARK Invest, Fidelity, Morgan Stanley, Sandisk, T. Rowe Price, Tether Investments and XTX Markets.

Tenstorrent says it will use the money to deploy its Galaxy servers at scale, serve open-weight models for LLMs and video generation, and advance the AI and CPU designs it licenses to chipmakers such as Samsung and LG. TechCrunch notes the valuation more than doubled from $2.6 billion in late 2024, and that the company reportedly drew acquisition interest from Intel and Qualcomm earlier this year. CEO Jim Keller: “The future of compute is open, scalable, and owned by the people building on it.”

Why it matters: Investors are backing Nvidia alternatives built around serving open-weight models. More competition in inference hardware could mean more choice and lower prices for running open models.

Sources: Tenstorrent newsroom · TechCrunch · Axios

---

What is doubleAI’s Maxwell?

Short answer: doubleAI, led by former Mobileye CEO Amnon Shashua, introduced a new category it calls Code Execution Models (CEMs). Its first, Maxwell, is in early access at heymaxwell.ai. Instead of writing code, Maxwell watches a running program and rewrites the machine instructions underneath it to fit the exact hardware and data. The source code stays unchanged, and it works on both CPUs and GPUs.

doubleAI says every optimization is formally proven to leave the program’s behavior unchanged. It reports 38x+ faster data pipelines and 4.5x+ faster AI inference kernels on real-world workloads. Calcalist notes these figures haven’t been independently verified. The company also announced a $200 million Series A co-led by Lightspeed and Bessemer, with Dell Technologies Capital, NVIDIA and others; Calcalist reports a valuation above $1 billion.

Why it matters: Compute is scarce, so a tool that makes existing hardware do more without code changes could cut AI bills. It complements, rather than replaces, coding assistants like GitHub Copilot and Cursor. Treat the speedups as claims until outside tests appear.

Sources: doubleAI (Business Wire) · Maxwell · Calcalist

---

What should builders watch next?

1. Federal vendors: guidance on what counts as a product “developed using DeepSeek implementations,” and whether other Chinese models get similar orders.
2. Agent security: expect more reports like NorForge’s. Lock down public admin panels and give your own agents separate, least-privilege identities.
3. The price of speed: MiniMax M3.1 Flash and Mercury 3 both target cheap, fast agent work. Watch for independent benchmarks and MiniMax’s promised open weights.
4. Compute efficiency: whether Tenstorrent’s servers and doubleAI’s Maxwell deliver lower inference costs outside company-run tests.

For tool deep-dives, browse the AI tools directory or the Models page. This roundup summarizes public announcements and reporting; it is not hands-on product testing.

Published by Andrew for AI Tools Optimizer.