Franz Bettag builds and operates his own platform for AI-assisted security research. It combines specialist agents, model selection for each workflow phase and fresh verification of findings. His security work has public recognition: #2 VDP researcher in Germany on HackerOne, January–March 2026.
AI can examine large codebases and attack surfaces. It can also produce plausible but false claims. A sound report needs an authorized scope, a reproducible vulnerability and demonstrated impact. Research, independent checking and documentation place different demands on a model.
Franz Bettag designs, builds and operates the platform for his own bug bounty work: backend and CLI, agent tools, phase prompts, model profiles and NixOS configuration. He defines the workflow and assesses its results through his work as a security researcher.
A Go backend stores campaigns, findings, submissions and runs centrally. Workers access its API through a shared CLI. NixOS defines the infrastructure. Each run records its model profile, start and end times, and status; artifacts such as logs and screenshots carry a run reference and SHA-256 checksum.
The inspected configuration includes OpenAI GPT/Codex, Anthropic Claude, GLM, Kimi, MiniMax and DeepSeek models. The launcher allows a separate model profile for each phase, so research and verification can use different models. Selection is explicit; this does not constitute an automatic ranking of the best model for each task.
Specialist agents investigate the selected scope. Platform-specific prompts account for the relevant reporting process. Findings and evidence are stored for subsequent phases.
Phase prompts require fresh, practical reproduction. A finding is classified as reportable, needs-review or false-positive. Ambiguous or unstable results should return for review; a model’s claim alone does not establish evidence.
Verified findings are examined for connected impact and prepared as draft reports. Instructions exclude purely theoretical chains and claims outside the authorized scope.
QA instructions require reproducible evidence, appropriate attachments and an assessment supported by demonstrated impact. Drafts receive ready or needs-revision status. Feedback feeds into further work.
The public event listing confirms second place for January–March 2026. The historical ranking screenshot shows fbettag with a 7.00 signal score and 21 reputation points in that view. This is a Vulnerability Disclosure Program ranking, rather than a global rank or a leaderboard of bounty payouts.
This work provides experience with model behaviour across long tool workflows, verification and conflicting results. For your use case, we turn that experience into concrete test cases, verifiable quality criteria and cost per accepted result. Your test set determines which model wins.
Architecture, phase instructions and model configuration were checked against the project’s own source code on 7 September 2026. This snapshot is separate from the historical ranking. No measured savings rate, overall report acceptance rate or causal contribution of individual models to that ranking is claimed here.
For clients, this becomes a scoped penetration-testing engagement: authorized targets, personally reviewed findings and coordination with the operations team. Monthly, weekly or daily cycles match your rate of change. We operate the models; clients receive no model access.
