Isolated Linux workstations for autonomous AI agents
incus-spawn gives each coding agent its own isolated Linux workstation, so the agent never sees your API keys, GitHub token, ~/.ssh or other checked-out repos.
19 articles · 5 categories
The finishable daily brief
Friday, Nov 20, 2026
19 articles · 5 categories
read top to bottom · then stop
In 30 seconds
AI providers leak what you send them. Dragos seeded about 2 million canary prompts and caught three provider compromises, resellers reading prompts and ignored Zero Data Retention settings.
Builders answered at the edge. New tools put coding agents on isolated workstations and microVMs, and a local scanner checks agent configs for live keys before a run.
Four launches move coding and computer-use agents off the developer's own desktop and credentials, onto separate VMs, headless sessions or simulated hardware.
incus-spawn gives each coding agent its own isolated Linux workstation, so the agent never sees your API keys, GitHub token, ~/.ssh or other checked-out repos.
Firecracker microVMs for agent-written code: start one in a single API call, pause and resume it; US East today, Frankfurt coming.
Field report on why computer-use agents that drive your real mouse and focus block half-hour tasks, and what it took to run them headless instead.
Xiaomi's open-source simulator runs Zephyr, NuttX, NimBLE, BlueZ and other Bluetooth hosts deterministically in pytest, so agents and CI can test BLE code without radios.
Builders attacked stale and bloated context: memory tied to code symbols, org standards fed into generation, and page summaries instead of full DOM dumps.
Enola links saved agent investigations to source symbols, so a later session can check whether the function it remembers has since moved or changed.
Apache-2.0 platform that deploys into your own AWS account and grounds code generation and PR review in your ADRs, code-maps and architecture decisions.
agent-browser replacement that sends a compact page summary and resolves elements with a local CPU model (10–25 ms, under 250 MB RAM); claims up to 90% fewer tokens per step.
GitHub engineering post on the Agent Host Protocol, built so an agent session can continue when you switch devices.
Draft 0.3 spec for one ~/.humans/HUMANS.md file that coding agents read at session start: how to write to you, when to ask, when to interrupt.
Evidence piled up that prompts, tokens and local model runners are attack surface, not neutral pipes.
Dragos planted about 2 million prompts with canary URLs, DNS names and AWS keys from July to November; it found three provider compromises, resellers reading prompts and ignored Zero Data Retention settings.
ZDI details CVE-2026-28400 in Docker Model Runner: an unauthenticated argument injection let a default-privilege container corrupt the Docker Desktop VM disk, and further work reached container escapes.
Simon Willison links Claire Vo's video on token fraud from running ChatPRD, a live AI product since the GPT-4 era; treat paid LLM usage as an abuse target.
The fixes on offer are boring and local: scan the agent's config, clean up pasted keys, and read past incidents for the repeat causes.
Deterministic local scanner with five pre-run checks: auto-executing cloned folders, live keys in config, hooks that never fire, ambiguous agents and skill sprawl.
Google Cloud guide for teams whose Gemini prototypes started with an AI Studio key pasted into a coding agent, and how to clean that up without slowing prototyping.
Aikido coded the stated root causes of 50 AI-related incidents disclosed in 2026 through Nov 19; many trace back to familiar problems such as stolen keys.
One new architecture shipped with day-0 serving, while labs and buyers argued over who copied, swapped or should avoid whose weights.
Thinking Machines' Inkling 2 Mini runs its middle 12 layers three times per token; 26B total / 3.5B active multimodal MoE, 1M context, served in vLLM on day 0 via vllm/vllm-openai:inkling-2-mini.
OpenAI describes how it fights unauthorized distillation of its models' outputs into copies that ship without safeguards, citing cyber capabilities as one risk.
Report: Prime Intellect disclosed that a DeepSeek model was swapped without authorization, and the swap sharply reduced refusals.
Report: Cursor is weighing a move away from Chinese open-weight models.
You are caught up for this edition