AIPwnRead research

AI security research for systems that act

Researching trust boundaries across agents, runtimes, and physical AI.

Read featured researchSubscribe

Independent · disclosure-first · agents to physical AI

Selected research

A curated snapshot of public work from AIPwn Research.

Browse publication feedresearch.aipwn.org/feed
  1. The Claude Code Fingerprinting Incident: Trust Boundaries and Defenses for Local AI AgentsLocal agents / trust boundaries
  2. AIBounty #000: 100 Days to PWN AI — The 2026 ResetResearch program
  3. AIPwn ·100 Days to PWN AIResearch log
  4. [paper] Prompt Injection 2.0 — The Hybrid AI ThreatPrompt injection
  5. [paper] Hacking the Hive Mind: How Multi-Agent LLMs Get JailbrokenMulti-agent security

Research scope

  • Prompt injectionModel layer

    Hijack models via untrusted input.

  • Agent abuseTool layer

    Tools turned against their owners.

  • Exposed systemsRuntime

    Open ports, panels, runtimes.

  • Physical AIReal world

    Robots acting on bad input.

Browse all research

Disclosure before publication

We coordinate with affected owners before publishing details.

Read disclosure policy