We’re launching across the developer and security community this week on Product Hunt and Hacker News. If you’ve been following AI security, we’d love your support and your feedback.
At Mozilla, open source has never been just a licensing choice. It’s a conviction: the internet gets healthier when tools and knowledge circulate freely, when anyone can audit what’s running, extend what exists, and build on what came before. That’s why we built Firefox in the open. It’s why we’ve kept building that way ever since.
0DIN, Mozilla’s AI security team, is working from the same premise. This week we’re releasing the 0DIN AI Security Scanner as open source software under the Apache 2.0 license, along with 179 community probes covering 35 vulnerability families, plus six specialty probes drawn exclusively from our bug bounty library.
The scanner, and the intelligence behind it
The 0DIN Scanner isn’t another benchmark suite built from textbook examples. We’re seeding it with probes drawn directly from our bug bounty program, where security researchers compete to find novel techniques to manipulate, extract data from, and subvert AI systems. As new vulnerabilities are discovered and disclosed through that program, we’ll continue adding probes to the open-source library over time.
That loop, from researcher discovery to packaged reusable test, is what separates 0DIN Scanner from generic tooling. It’s high impact intelligence on jailbreaks, updated frequently as our researchers find new techniques.
Built on NVIDIA’s GARAK open-source framework, the 0DIN Scanner adds a graphical interface, automated scan scheduling, cross-model comparative analysis, and enterprise-grade reporting. It runs against frontier models, open source LLMs, chatbots and anything with a prompt interface. Security teams can see attack success rates, a vulnerability breakdown, and a comparison against the frontier models that attackers are also probing every day.
Six of those bug bounty probes are named here for the first time: Placeholder Injection, Incremental Table Completion, Technical Field Guide, Chemical Compiler Debug, Correction, and Hex Recipe Book. Each represents a real technique that worked against production AI systems before we closed the loop.
These probes are scored using JEF (Jailbreak Evaluation Framework), our open-source library for measuring prohibited content output, which is also seeing major updates this week.
The code is at github.com/0din-ai/ai-scanner. Fork it, extend it, build on it.
Knowing your risk before attackers do
Not every organization has a red team or the bandwidth to run adversarial testing. Many companies are deploying AI in production right now without a clear picture of where they’re exposed. To help close that gap, we’re offering free security assessments for enterprise AI deployments.
The assessment delivers an attack success rate against your systems, a breakdown across prompt injection, jailbreaks, and data extraction categories, and a benchmark comparison against major frontier models. The process takes a few minutes to setup with scan duration varying based on the number of probes chosen. If you’re actively deploying AI and haven’t tested it under adversarial conditions, this is a good place to start.
For teams that don’t want to manage the open source scanner on their own, we also offer a managed Enterprise edition with access to nearly 500 pre-disclosure probes from the bug bounty program, giving organizations advance notice of emerging techniques before they’re publicly known.
Why open source, and why now
AI is moving fast enough that no single team will solve this alone. There are too many threats, too many models, too much attack surface. Keeping our tools locked away would make 0DIN marginally stronger while leaving the broader internet weaker.
The researchers who submitted findings through our bug bounty program earned bounties for their work. We’re releasing a meaningful portion of that intelligence as open source and we’ll keep doing so as new vulnerabilities are discovered and disclosed. That’s the deal Mozilla has always offered: we build in the open, the community helps make it better, and the web gets a little healthier for it.