One bug-reporting system to replace dozens of broken forms

Researchers who discover a flaw in a language model or deployed AI system currently face a maze: no single address, no uniform format, no guarantee the information reaches whoever needs to fix it. The FLARE-AI paper published on arXiv presents an open-source system built for exactly this problem — a single platform for creating a standardized incident report and distributing it automatically to multiple developers, security coordinators, and incident registries at once.

The problem: a fragmented ecosystem

The team surveyed 12 existing reporting systems — from model developers, cyber bodies, and vulnerability aggregators — and identified five recurring failures: difficulty finding the right address, ambiguous scope definition, partial information collection, lack of coordination among report recipients, and missing guidance for strict-liability cases. The practical result: good-faith reporters fill out different forms for the same flaw, and recipients are left with non-standardized information that is hard to triage.

Design process: 49 experts, 32 organizations

To avoid swapping one guess for another, the researchers gathered feedback from 49 experts representing 32 organizations — developers, security researchers, and ecosystem coordinators. Insights from that round shaped FLARE-AI's architecture: conditional logic that guides the reporter only to relevant fields, upfront classification of flaw type, and a machine-readable JSON export in a uniform format that can be sent to multiple recipients from a single submission.

How it works in practice

The reporter fills out a single form that adapts in real time: select a flaw category, and the system surfaces only the fields required for that type — for example, details of a harmful prompt, unexpected output, or personal-data leakage. On submission, FLARE-AI produces a standardized JSON package and enables optional distribution to a predefined recipient list: the relevant developer, a coordination body such as a CERT, and a public incident registry. All open source, with no single-vendor dependency.

Seeking an organizational home

The researchers state explicitly that they lack the bandwidth to maintain the project long-term as individuals, and are looking for an independent organization to take full ownership. This is critical: without a body committed to hosting, security updates, and version management, the system risks becoming yet another abandoned GitHub project — exactly the kind of fragmentation it was designed to solve.

Why it matters

The real value lies not in the technology itself but in the convention it tries to establish: a single reporting format, a shared language for developers and coordinators, and a clear path from discovery to fix. If it succeeds, FLARE-AI will turn "where do I report this?" from an open question into a technical default — and that, more than any benchmark, is what advances AI safety at scale.