Stanford's ARTEMIS turns Codex into a red-team swarm

What it does ARTEMIS is an autonomous agent from the Stanford Trinity project that automates vulnerability discovery. It wraps a fork of OpenAI Codex in a supervisory layer that spawns and manages multiple subagent instances to probe code for flaws. The README frames it as a hands-off red teaming engine, though it offers little detail on methodology beyond the multi-agent architecture.

The interesting bit Rather than building a scanner from scratch, the project grafts a multi-agent orchestration layer onto OpenAI Codex, turning a coding assistant into a security-focused swarm. The supervisor delegates tasks to spawned codex-rs instances, effectively treating vulnerability hunting as a distributed coding challenge.

Key highlights

• Built on a fork of OpenAI Codex, extending it with a Python-based supervisor.
• Spawns Rust-based Codex subagents (codex-rs) to perform delegated security tasks.
• Evaluates itself against CTF-style challenges in benchmark mode.
• Requires external LLM API keys (OpenAI or OpenRouter) to function.
• Licensed under Apache-2.0.

Caveats

• The README is overwhelmingly quickstart and setup instructions; actual technical depth on discovery capabilities or architecture is thin.
• It is unclear how the supervisor triages findings or avoids duplicate work, since the docs focus on installation rather than operational logic.
• Requires paid third-party API access and specific model configuration before it can run a single test.

Verdict Security researchers experimenting with LLM-driven red teaming should keep an eye on this. If you need a standalone scanner that works out of the box without API keys, look elsewhere.