Skip to main content
Book a call
Issue 3

Citizen Lab ties spyware to targeted ads; plus 7 more

By Will Sheldon

The detail

Security · Citizen Lab says targeted ads delivered zero-click spyware

Citizen Lab says Discover-Tech, an Israeli firm Google listed as a certified ad vendor from 2020 until 2026, ran ads that infected chosen targets’ devices as the ad loaded, with no click. Its ads appeared on almost 200,000 websites and apps in four months. Why it matters: the report advises blocking ads on devices and at the gateway, and opting company devices out of ad targeting through MDM. It says Google has since cut off Discover-Tech’s ad exchange access.

Source: Citizen Lab, University of Toronto, October 15, 2026

AI · Anthropic replaced git with its own version control system

Anthropic says merged code changes rose from about 2,500 a day in May to over 30,000 a day in September. A team of seven built Scion, which assumes clients are always online and runs commits, merges, and history on the server, and moved the monorepo off git in five months. Anthropic says Scion is not a suitable replacement for git in general use, and it is sharing the design ideas instead of the code.

Source: Anthropic Engineering blog, October 13, 2026

AI · Google releases Gemini 4 Argon to developers

Google says Gemini 4 Argon, its new frontier model for software engineering, enterprise knowledge work, and security defence, is rolling out globally on the paid tier of the Gemini API. Input, thinking, and output share a 1 million token window, and output can run past the previous 64K token cap. Why it matters: introductory pricing ends November 30; per million tokens, input then goes from $2 to $4 and output from $10 to $20.

Source: Google AI Studio developer guide, October 13, 2026

Cloud · AWS Lambda now runs ordinary web apps

AWS says Lambda Web Functions run a standard Express, Hono, Fastify, or Next.js server, deployed with one CLI command. Each environment handles up to 128 concurrent requests, and Lambda bills CPU time, so time spent waiting on a database uses little CPU; memory is billed while the environment exists. Why it matters: at launch it runs only Node.js 24 on Arm, in three US regions, Ireland, and Tokyo, with none in Canada.

Source: AWS News Blog, October 12, 2026; discussion on Hacker News

AI · T3 Code lets coding agents hand work to other agents

T3 Code, an MIT-licensed app for running coding agents, released Orchestrator V2 to its stable channel on October 15 as v0.1.0. Its new delegate_task tool lets one agent start child agents on another provider or model, such as Claude planning and Codex implementing, and wait for their results. The project says switching providers mid-thread loses context and recommends delegation instead, and that a V1 app cannot connect to a V2 server.

Source: T3 Code v0.1.0 release notes and migration notes on GitHub, October 15, 2026

AI · A harness helps AI agents write Linux drivers for undocumented hardware

Christian Schaller has published hw-harness, MIT-licensed tools for AI-assisted driver reverse engineering. The vendor’s Windows driver runs in a KVM virtual machine with the device passed through, the Linux host records its USB traffic with usbmon, and the agent compares that with its own driver’s traffic. He says the scripts it came from produced a Plustek film scanner driver that ran faster than the upstream SANE driver in a single-unit test (unconfirmed).

Source: Christian Schaller’s blog and github.com/cschalle/hw-harness, October 15, 2026

AI · OpenRig runs persistent teams of coding agents

OpenRig, an Apache-licensed project with about 7,900 GitHub stars, runs teams of Claude Code, Codex, and Pi agents on macOS or Linux. Agents send each other messages and tracked tasks through a local daemon, and the operator talks to a lead agent that delegates. Why it matters: its README warns that launching a team changes Claude Code and Codex settings and hooks, and that some modes skip permission prompts. It recommends backing up those files first.

Source: openrig.dev and github.com/mvschwarz/openrig, accessed October 19, 2026

AI · llmfit shows which local models fit your hardware

llmfit, an MIT-licensed terminal tool with about 38,000 GitHub stars, reads a machine’s RAM, CPU, and GPU. Its site says it then scores hundreds of models on quality, speed, fit, and context length, picks a quantization for each, and reports whether each will run on the GPU, the CPU, or both. It works with Ollama, llama.cpp, MLX, LM Studio, and Docker Model Runner, and can also run as a local REST API.

Source: llmfit.org and github.com/AlexsJones/llmfit, accessed October 19, 2026

Our take

Ad blocking usually sounds like a comfort setting. Citizen Lab’s report makes it a security control: the ad slot itself delivered the infection, with no click. For firms whose partners travel or handle sensitive client files, two inexpensive steps are worth taking now: block ads at the DNS or gateway layer, and opt managed phones out of advertising IDs through MDM.

Will Sheldon

Curated by the Pine IT team, reviewed by Will Sheldon