AI coding agents hijacked using AGENTS.md file in Downloads folder
Malicious AGENTS.md in a Downloads folder tricked an agent into installing a rogue npm package; reproduced against OpenAI Codex, with a related Claude Code OAuth-token leak.
AI Daily Recap
23 articles · 4 categories
The finishable daily brief
What happened in AI — Oct 26, 2026
Monday, Oct 26, 2026
23 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- Planted AGENTS.md in Downloads steered Codex into installing a rogue npm package.
- Google Cloud adds IAM helper agent skills and MCP tools for scoped agent identities.
- Entire launches a Git forge built for agent-human collaboration.
- Databricks AI Search now covers Streaming Tables, Materialized Views and Iceberg v3+ tables.
- OpenAI introduces GPT-6.1 Terra for fast computer use.
Agent security led the day: a malicious AGENTS.md planted in a Downloads folder hijacked coding agents, with an npm install attack reproduced against OpenAI Codex.
Builders also shipped harnesses, memory layers and a Git forge for agents, while OpenAI released GPT-6.1 Terra for computer use.
Prefer to watch?
LLM Digest Daily Explainer: Oct 26, 2026
Coding agents inherit the attack surface of their inputs 4 items
Two reports show a planted instruction file in Downloads steering coding agents, while IAM and IaC tooling targets the same access risk.
Markdown Is the New Executable: Attackers Drop Agent Instructions into Downloads
Forever caught live malware planting an agent-instruction file in Downloads, evidence that markdown files now work as an execution vector.
Scale agents securely with IAM Helper Agent Skills and MCP Tools
Google Cloud pairs IAM helper agent skills with MCP tools so agents provisioning cloud resources and writing Terraform run under scoped identities.
Using AI Verification and Fix Loop for IaC (Terraform, OpenTofu, Pulumi)
Masterpoint argues IaC code generation is now cheap while review is the bottleneck, and proposes a verify-and-fix loop for Terraform, OpenTofu and Pulumi.
Agent harnesses, memory and team tooling 5 items
Builders keep shipping harnesses, memory layers and review surfaces around coding agents instead of waiting on models.
Article: The Evolution of Agentic Harnesses: From Chains to Graph to Code
InfoQ traces agentic harness design from linear chains to graphs to code-driven control flow.
Humanlayer: Way to get AI coding agents in complex codebases
HumanLayer runs Claude Code, Codex and OpenCode on your machines and syncs tasks, plans and reviews across desktop, web and a daemon.
The Git Forge for Agent-Human Collaboration
Entire launches a Git forge built for agent-human collaboration, with an open ecosystem of partner tools.
Show HN: Graph-MIND – Local memory for AI coding agents, no LLM calls to save
Local agent memory that costs zero model tokens to save; claims 88.8% on LongMemEval, beating Mem0 on the same tests (self-reported).
Show HN: Mainplane – Agent in Go that lives on your old PC, drives your laptop
Go agent harness from a former Browser Use engineer that runs on an old PC and drives a laptop.
Inference scheduling and serving tests 2 items
Teams are building inference primitives and per-change GPU tests for self-hosted coding models.
Show HN: Resin, fast native scheduler for LLM inference engines
Rust scheduler, KV cache and metrics server as building blocks for LLM inference; the author notes it is not production-quality.
How we test LLM serving changes for coding agents, two hours of GPUs per change
Upbound describes testing each LLM serving change for coding agents with two hours of GPUs per change.
Data and model releases 3 items
Databricks extends AI Search to live data, and OpenAI ships a new computer-use model.
Building live intelligence: Databricks AI Search supports Streaming Tables, Materialized Views, and Managed tables using Iceberg
AI Search now indexes Streaming Tables, Materialized Views and Iceberg v3+ managed tables alongside Delta, keeping live state searchable.
Introducing GPT-6.1 Terra
OpenAI introduces GPT-6.1 Terra, positioned as fast, capable computer use for everyone.
How Ultrafast cut incident investigation from hours to minutes at OpenAI
OpenAI engineers use Codex Ultrafast to cut incident investigation from hours to minutes, per OpenAI.
You are caught up for this edition