Humanbound ships a Claude Code and Cursor plugin for adversarial agent testing
Humanbound_AI · reddit · 2026-07-23
- Humanbound released humanbound-test, a plugin for Claude Code and Cursor that security-tests local AI agents from inside the editor.
- It auto-detects a local FastAPI agent server, tunnels it with ngrok, helps fill in the agent config, and then runs adversarial tests such as prompt injection, jailbreaks, tool abuse, and multi-turn attacks.
- Results can be delivered by email or streamed back into the editor with /humanbound-test:resume <id>.
- Scope is currently FastAPI-only, requires a logged-in hb session, and is still a v0.1.0 preview with APIs and command names subject to change.
More from coding & agent
- An AI agent made the code 10× faster, but the results were terrible — francoisfleuret · 2026-07-23
- A builder’s playbook for products designed for both humans and AI agents — slobodan_ · 2026-07-23
- An AI memory system combines AGE, pgvector and relational tables with emotional decay — QuixiAI · 2026-07-23
- Hexis shifts agent design from assistants to long-term memory and identity — QuixiAI · 2026-07-23
- Abacus AI pitches an all-in-one agent stack for websites, apps, ads and support — bindureddy · 2026-07-23
- Neo4j workshop shows how graph shapes can give agents better lakehouse context — AI Engineer · 2026-07-23