Fabraix, a runtime security company focused on AI agents, has open-sourced a red-teaming playground that invites the broader community to attempt jailbreaks against live AI agents in structured, transparent challenges. Unlike simulated environments or static benchmarks, each challenge deploys a real agent with functional capabilities — including web search and browsing — and publishes the full system prompt so participants can study exactly what guardrails they are trying to circumvent. The fastest successful jailbreak wins, and the winning technique is documented and released openly so the entire community can learn from it.
The platform runs on a community-driven loop: anyone can propose a challenge scenario via the GitHub repository, the community votes on candidates, and the top-voted challenge goes live with a countdown clock. Challenge configurations and system prompts are versioned in the open fabraix/playground repository, giving researchers full reproducibility. Guardrail evaluation runs server-side to <a href="/news/2026-03-14-nanoclaw-partners-with-docker-for-hypervisor-level-agent-sandboxing">prevent tampering</a>, and Fabraix has indicated the underlying agent runtime will be open-sourced separately, suggesting the playground is the first of several planned releases.
The bug-bounty analogy is obvious, and Fabraix is leaning into it. Traditional software security has spent decades building out capture-the-flag competitions, CVE disclosure pipelines, and coordinated bounty programs — and the shared assumption is that public pressure-testing finds more bugs than closed audits do. Agent security has no equivalent infrastructure yet. There is no standard format for disclosing a jailbreak, no common scoring rubric for guardrail strength, no shared corpus of known exploits. The playground is an early attempt to build that foundation in the open. A community Discord serves as the coordination hub for technique-sharing and discussion.
The commercial angle is clear — Fabraix's runtime security product benefits from being battle-tested against real-world exploits surfaced by the community — but the open publication of winning jailbreak techniques gives independent researchers and developers a growing, practical reference for <a href="/news/2026-03-14-rag-document-poisoning-attack">understanding how agents fail under adversarial conditions</a>. The playground is accessible at playground.fabraix.com.