An independent lab for AI agent security.
secagentlabs studies how autonomous AI agents fail under attack — and how to build them so they don’t. We work in the open: reproducible attacks, honest defenses, citable sources.
What we publish
Offensive research (attack techniques and proofs of concept) and defensive engineering (controls tested against those attacks, including their bypasses). No vendor pitches.
Editorial standards
Every piece is human-reviewed and edited by the secagentlabs Research editorial team before publication. We don’t invent clients, numbers or incidents — claims map to public CVEs, papers and code, and we correct mistakes in the open.
Vendor-neutral
There is no sponsorship in the research. When we name a tool, it’s because it broke or it held — not because someone paid for the mention.
Responsible disclosure
Found a vulnerability in an agent, framework or MCP server? We coordinate disclosure with affected parties and credit reporters. Reach us in English, Polish or German.
contact@secagentlabs.com →