Inside · A2Inception's Mercury 3 scales diffusion language models to a million-token context Inside · A3Benjamin Congdon makes the case for agents that watch every release
Today’s forecastA Federal Register front moves in over DeepSeek, with hidden instructions drifting through help pages and a record Linux kernel breaking through late Sunday.

The Daily Archit

“All the tech that’s fit to print”Late City Edition
Vol. I · No. 3 Monday, October 19, 2026 Price: one HTTP request
The Newsstand · all editions
Image via Federal Register
POLICY

DHS orders DeepSeek out of every federal civilian agency

A supply-chain order signed in July and made public on Monday also covers third-party products built on DeepSeek technology.

By the Daily Archit news desk Federal Register

Homeland Security has ordered civilian agencies to stop buying DeepSeek products and remove them from their systems now. The order covers affiliates and anything built on DeepSeek models, and agencies have 60 days to ask CISA for exceptions.

Private use is untouched, but vendors who quietly swapped in DeepSeek weights now have a compliance problem.

Full story at Federal Register →
KERNEL

Linux 7.3 lands as the busiest kernel release on record

Linus Torvalds released Linux 7.3 on Sunday: 18,765 commits from 3,021 developers, 807 of them first-timers. It completes sched_ext subschedulers and adds an fchroot() system call, per-CPU global data for BPF, VM working-set tracking and failfs, a filesystem that fails every operation, which is handy for testing. The 7.4 merge window is open.

LWN.net
Image via MiniMax
MODELS

MiniMax M3.1 Flash targets agent work at 15 cents per million input tokens

MiniMax's new Flash model costs $0.15 per million input tokens and $0.60 per million output. The company says it runs above 170 tokens a second and scores 73.7% on DeepSWE and 61.7 on OSWorld V2. All the benchmark figures are MiniMax's own. Weights are promised "later", with no licence named yet.

MiniMax
Image via TechCrunch
AGENT SECURITY

Microsoft tells enterprises to treat AI agents like insider threats

Microsoft launched Guardian Agents to watch the agents a company runs. Defender can now spot about 60 kinds of agents that staff bring in, Claude Code and Codex among them. New controls scope permissions to the task and require human approval for risky actions. Microsoft also invests in the labs whose models it now says can't be trusted.

TechCrunch

From the Newsstand see every edition →

The Daily Archit No. 9 · Sunday, October 25, 2026 OPEN-WEIGHT MODELS Moonshot's Kimi K3.1 Pro pulls an open-weight model to within a point of the frontier OpenAI takes frontier model training off the live internetAnthropic hands Agent Skills to the Linux Foundation's Agentic AI FoundationBrotli decoder bug grows from Node2Shell into HTTP2Shell, reaching Python, Bun, Envoy and curl The Daily Archit No. 8 · Saturday, October 24, 2026 LANGUAGES C++26 fails its ISO ballot, leaving a finished standard in limbo Mozilla pulls trust from an eMudhra root as the CA blames social media for missed deadlinesSam Altman says defenders can no longer patch their way out of AI-found bugsAsked to pick midterm candidates, Grok goes Republican and GPT goes Democrat The Daily Archit No. 7 · Friday, October 23, 2026 AI & MEDICINE Microsoft's diagnostic agent beats generalist doctors in a peer-reviewed Science study Go hard-codes distrust of a set of compromised intermediate certificatesCanada's lawful-access bill would let a minister order backdoors into messaging servicesGermany's AfD confirms a 281 GB leak of internal email after its host was hacked The Daily Archit No. 6 · Thursday, October 22, 2026 SUPPLY CHAIN Poisoned LiteLLM release on PyPI stole the keys to the AI stack Railway tells every user to rotate secrets after an attacker reached stored environment variablesAnthropic and Theorem prove 239 fuzz targets crash-free, down to the compiled binaryeMudhra quits the global TLS certificate business after its CA keys proved trivially weak