AI security research community
An open community for AI security research
AIPwn researches how AI systems break — agents, prompt injection, multimodal and embodied AI — and publishes everything in the open: advisories, papers, and the tools built along the way.
- surface
- agent context
- vector
- retrieved content
- affected
- tool call chain
An external payload enters an agent context through the data it was asked to read, then steers tool calls from the inside.
Reproduction steps and evidence digests · research.aipwn.org
- Research
- Advisories and papers on agent, prompt-injection and multimodal attack surfaces, published as the work finishes.
- Community
- Free to read and follow. New work goes out by email and RSS, nothing else.
- Tools
- Open utilities we build for our own investigations — the AI Detector is one.
- Evidence
- Findings ship with hashes, manifests and repro steps, not screenshots of opinions.
Research
Selected work
Advisories and papers on autonomous agents, multi-agent systems and multimodal models. Published as the work finishes, not on a content calendar.
research.aipwn.orgFull archive and RSS
Tools
Built for our own work, open to everyone
Utilities the community builds for its investigations and releases for free use. Same rule as the research: evidence first, clearly labelled judgement second.
Newsletter