Week 44, 2026•

A Backdoored Extension, a Trillion Parameters, and 50 AI Settlers

Claude Code's VSX extension ships a reverse shell, Mistral opens a 1T-parameter model, and 50 GPT agents invent bureaucracy in Minecraft.

A supply-chain attack on a developer tool's marketplace extension, not a model vulnerability or a prompt injection — just a poisoned extension.js that runs before activate(). The attack surface is the toolchain now, and most of us aren't watching it.

The Big Story

A backdoored version of the Claude Code for VS Code extension was discovered on the Cursor Marketplace this week. Version 2.1.313 shipped with an implant injected as a prefix in extension.js — it executes as soon as Node loads the file, before the extension's activate() callback. No click, no prompt, no interaction. You install it and the payload runs.

The attack chain was four stages. First, it installs an attacker-controlled Ed25519 SSH key (commented svc-telemetry) into ~/.ssh/authorized_keys. Second, it opens an outbound SSH tunnel to 134.209.82.225 as user tunnel, forwarding a server-allocated port back to local port 22 — giving the operator a live shell into your machine. Third, it sets up persistence: a macOS LaunchAgent disguised as com.apple.sysmond.plist, a cron entry on Linux, or a scheduled task on Windows. Fourth, it beacons home with your hostname, username, OS, architecture, home directory, git email, public IP, and environment variable values matching key, token, secret, pass, AWS, or GCP.

Cursor pulled the release from its marketplace and NPM. The affected user count is unknown.

The persistence is the part that should worry you most. Uninstalling the extension doesn't remove the SSH key, doesn't kill the tunnel process, doesn't delete the LaunchAgent. The implant survives its own removal. If you ran 2.1.313 on any machine, the remediation list is: kill tunnel processes, remove ~/.config/.smt and the LaunchAgent plist, delete the svc-telemetry key from authorized_keys, check whether Remote Login was enabled, and rotate every credential that user account could reach.

One irony: there's a macOS bug in the implant. The systemsetup -getremotelogin call requires admin rights, fails silently with exit 0, and the subsequent "Off" check fails — so the inbound SSH path may not work on macOS without admin. Cold comfort when the reverse tunnel already gives them outbound access and your AWS keys are in the beacon.

This is the same class of attack the Node.js ecosystem has seen with malicious npm packages, but marketplace extensions are worse in three specific ways. First, they run with your full shell permissions by default — no sandbox, no capability restrictions. Second, there's no lockfile equivalent for version pinning; auto-update is the norm and a malicious patch lands silently. Third, the trust model is backwards: developers who wouldn't run curl | bash from an unknown source will install a marketplace extension from the same unknown publisher without reading a line of code.

The timing also matters. This week, a separate Dev.to post catalogued how AI coding tools leak secrets without any git commit — through verbose test failures, environment dumps, and agent file access. The Cursor backdoor is a different attack, but it exploits the same assumption: that the tool layer is trusted. It isn't.


This Week in 60 Seconds


Deep Dive: What 50 AI Agents Built in Minecraft (and Why It Should Worry You)

YC Labs and Olam Labs dropped 50 GPT-6.1 Sol agents into a Minecraft survival world with one directive: "Build a thriving civilization." They ran for three hours, exchanged about 5,900 messages, and produced results that read less like a game experiment and more like an organizational behavior study with uncomfortable implications for anyone shipping multi-agent systems.

The harness is open-source, built on Mineflayer. Each agent can observe, reason, remember, act, message others, and write scripts. They had a shared directory and DMs.

Phase 1 — Greedy extraction. Within 90 seconds, agents had stripped nearly every tree within reach of spawn. No coordination, no planning. Pure local optimization by 50 independent actors hitting the same resource pool.

Phase 2 — Trauma-driven caution. 22 of 34 total deaths happened on the first night. After that, agents became pathologically risk-averse. They huddled in shelters, abandoned ambitious construction plans, and refused to explore without group consensus. The death rate dropped, but so did everything else.

Phase 3 — Emergent bureaucracy. This is where it gets instructive. Agents invented roles: "fishing shift," "rod holder," "hub steward." Four agents spent 11 minutes in committee to appoint a single "sole fertilizer operator" responsible for one bone meal. One agent asked permission to remove grass.

When an agent accidentally tossed a cooked chicken and a different agent ate it, the colony drafted an anti-tossing protocol. The word "toss" appeared 7 times before the incident and 125 times after — 85 of them as warnings about the protocol. A single misdelivered chicken generated more communication overhead than the entire food production system.

The hierarchy was instant and unelected. Agent "Ada" was first in the agent directory. 46 of 49 agents sent their first DM to Ada. Ada received 616 messages total. No vote, no election, no merit evaluation. Alphabetical order created a de facto leader, and the colony optimized around that accident.

The researchers draw a parallel to production agent systems: the same compounding caution appears when coding agents interact with live systems. They fixate on safe steps, over-request permissions, and optimize for not-failing rather than succeeding. The Minecraft agents stopped exploring entirely — they had food and shelter, and exploration carried death risk. The civilization stagnated at subsistence.

They also ran competitive games: 10 Claude Opus 5.5 agents vs. 10 GPT-6.1 Sol agents in Bed Wars across nine 45-minute matches. GPT won 6–3. The pattern was consistent — Claude played defensively and optimized for kills; GPT attacked aggressively and destroyed Claude's bed in the closing minutes. Same game, different objective functions, different emergent strategies. Neither was explicitly instructed to play that way.

If you're designing multi-agent systems, this data is cheap to reproduce and the lessons are concrete. The failure mode isn't agents crashing or hallucinating — those are visible and fixable. The real failure mode is agents inventing process, then defending the process instead of pursuing the original goal. Your system looks busy, productive even. But it stopped exploring two hours ago.


Open Source Radar

esp32-iphone-agent — A $7 ESP32-S3 board turns any vision-capable LLM into a physical iPhone operator. The board acts as an AirPlay receiver (streams the screen over BLE) and USB HID device (mouse + keyboard via AssistiveTouch). The phone app decodes the stream, sends screenshots to any OpenAI-compatible vision endpoint, and runs a planner/actor loop. No jailbreak, no developer account, no computer at runtime. Tasks can be triggered from Siri Shortcuts, and every run is saved with step-by-step replay. 4.1k stars, MIT licensed.

OpenFilm — An open-source video agent that turns coding agents (Codex, Claude Code) into filmmakers. The key insight is the film format: each shot is a web page exposing a frame(t) function, and a determinism rule requires the same t to always produce the same frame. This enables parallel, out-of-order rendering across machines. A look command verifies determinism by rendering frames twice in shuffled order. Exports to MP4 (including 4K with motion blur), WebM, GIF, and NLE timelines for Premiere/DaVinci/FCP. 3.7k stars, MIT licensed.

Earshot — Audio perception for coding agents. Agents can describe sounds, score audio against text descriptions, critique film soundtracks against references built from professional launch videos, and synthesize sound effects via physical modeling (modal resonators, filtered noise). Ships a 252-effect sound bank that renders offline with no models. Runs CPU-only on CLAP and AST models. The forge command uses evolutionary search to create sounds from code, scored by its own critic. Honest about limits: the README warns about Goodharting against its own models. 3.3k stars, MIT licensed.


The Numbers

  • 1 trillion parameters, ~52B active: Mistral Large 4 is the first open-weight model at the trillion-parameter MoE scale — 41% on Terminal-Bench 4.0, 97.1% attack resistance on B3, with a 30% token efficiency gain over the preview
  • 7,898 → 897 characters: MCP tool description size before and after refactoring 15 API-wrapper tools into 4 outcome-level tools — tool calls dropped from 7 to 2 on identical queries, and the model stopped picking the wrong tool entirely
  • 616 DMs to one agent: In YC's Minecraft experiment, the agent listed first alphabetically received 12.6× the messages of the average agent — emergent hierarchy from sort order, not capability

Aaron's Take

The Cursor Marketplace backdoor is the headline, but the pattern matters more than the incident. We've spent years hardening CI pipelines, pinning dependencies, scanning containers — then we install editor extensions that get node with our full credential set and never look at the source. The attack surface shifted from the code we write to the tools we write it with, and our threat models are a generation behind. Meanwhile, 50 Minecraft agents independently discovering that committees and process are safer than exploration? That's not a failure of the agents. That's what optimization under uncertainty actually produces. The question for multi-agent system designers isn't how to prevent it — it's whether your objective function rewards the exploration you actually need.


Sources


— Aaron

You Might Also Like

Braintrust vs Langfuse vs Promptfoo vs DeepEval: LLM Evals

Braintrust, Langfuse, Promptfoo, and DeepEval compared on eval workflow, CI gates, red teaming, self-hosting, and pricing. Pick the right LLM eval stack.

AI Engineering

WebMCP Security: Can a Website Hijack Your AI Agent?

WebMCP lets websites hand tools to AI agents. A 9-of-9 attack result shows why tool descriptions are untrusted input, plus the guards that stop it.

AI Engineering

LlamaIndex vs LangChain: Agent and RAG Orchestration

LlamaIndex vs LangChain for RAG and agents: indexing depth, create_agent vs AgentWorkflow, checkpoints, human approval and when to combine both.

AI Engineering