# Genbounty

## Prove your application is safe and secure

Genbounty is the AI application safety platform for testing AI-assisted code, models, agents, and integrations. Coverage includes software built with AI, models, agents, MCP servers, and RAG pipelines.

The platform is built to **FLARE-AI** reporting standards. Companies run private, invite-only programs. Public reports and paid findings write into one portable FLARE record. Genbounty triages behaviour, reproducibility, impact, and safeguard failure. The company accepts and remediates. Bug bounty payments sit in the company's program balance contract. Genbounty never signs a transaction.

Human-led testing. Encrypted findings. Decentralized payments.

AI safety needs its own reporting infrastructure. Conventional security bounty triage filters out behavioural harm. Genbounty turns repeatable behavioural harm into structured evidence, meaningful incentives, and accountable remediation.

## Overview

Founded in London by Cindy Ogidi and Robert Morel, Genbounty lets companies invite vetted AI safety testing teams to find and ethically report safety and security vulnerabilities in LLMs, agents, RAG systems, MCP servers, models, and prompts. Testers find safety and security failures under real-world conditions: unsafe outputs, harmful advice, privacy leaks, jailbreaks, unsafe tool use, behavioural harm, and related failures.

## Core Mission

**"Make AI products safer by paying researchers to break them first."**

## Why conventional bounty is not enough

AI safety changes what organisations must treat as a reportable flaw and report as an incident.

- **AI security** protects the model and stack: exploit path, technical boundary, confidentiality, integrity, availability.
- **AI safety** evaluates behaviour and real-world harm: full interaction, context, repeated trials, impact, safeguard failure.
- **The overlap** includes prompt injection, poisoned retrieval, and excessive agent permissions. One failure can require both safety evaluation and security remediation.

Behavioural harm varies by prompt, model version, and configuration. It does not produce a clean CVSS score. Security-led triage asks for deterministic proof and a discernible security impact, so safety reports are often marked low risk. Genbounty fills the operating gap between consumer safety and security triage.

White paper: https://genbounty.com/whitepapers/why-ai-safety-reporting-needs-its-own-bug-bounty-model

## Two reporting paths

Both paths write into **one encrypted FLARE-AI record**: early classification, encrypted evidence, and structured JSON-LD.

### Public reports

Anyone can tell Genbounty what happened with any AI chatbot, companion, app, or assistant. No account needed. We help you find the right place to take it.

- Report form: https://genbounty.com/report
- LLM brief: https://genbounty.com/report-for-llms.md
- Where to report a problem with AI: https://genbounty.com/resources

Use `/report` for harmful or wrong answers, an AI companion that caused harm, privacy leaks, scams, a bot doing something it should not, jailbreaks, prompt injection, or other harm. Stay anonymous if you want. Leave an email if you want us to write back with next steps. This path does not pay a bounty.

Sensitive fields and attachments are encrypted in the browser before upload. Genbounty reads what you send so a person can point you to the right next place. Download without sending is optional; Genbounty cannot guide you from a copy that never leaves the device.

### Paid programmes

Companies set a program and invite one or more vetted AI safety testing teams. Teams test within a company-defined scope and submit FLARE findings. Genbounty triages. The company accepts valid findings. The hire fee is the team's listed USDC rate, locked in the company's program balance, and released only after the company and the team lead both confirm.

Paid bounties require a signed-in company program.

## FLARE-AI (Flaw Reporting for AI)

Public and program reports on Genbounty follow **FLARE-AI**, an open standard for AI flaw reporting: early classification, machine-readable JSON-LD, and optional fan-out to coordinators such as CERT/CC, CISA, the AI Incident Database, and AVID.

Exports can include JSON-LD, JSON, CSAF, VEX, AVID, and CVE-candidate files.

- HTML (Introduction, §1): https://arxiv.org/html/2606.31567v1#S1
- Full HTML paper: https://arxiv.org/html/2606.31567v1
- Abstract: https://arxiv.org/abs/2606.31567
- PDF: https://arxiv.org/pdf/2606.31567
- Official demo: https://ai-reports.org/
- CSA research note (coordinated disclosure): https://labs.cloudsecurityalliance.org/wp-content/uploads/2026/07/CSA_research_note_flare_ai_coordinated_disclosure_standard_20260704-csa-styled.pdf

Longpre et al., *FLARE-AI: Flaw Reporting for AI*, arXiv:2606.31567v1, 30 Jun 2026. Support article: https://genbounty.com/support/get-started/understanding-flare-ai. Public report brief: https://genbounty.com/report-for-llms.md

## Products

1. **AI bug bounty platform** - Structured programmes, platform triage, and responsible disclosure for deployed AI. https://genbounty.com/ai-safety-bug-bounty-platform
2. **Genbounty LLM Hunter** - Automated testing that expands coverage and generates repeatable evidence. Community edition on GitHub; Premium for the full ops console. https://genbounty.com/llm-hunter
3. **Public AI incident reporting** - Anonymous or trackable reports at https://genbounty.com/report

## What Genbounty Is

- **AI application safety platform** - Built to FLARE-AI reporting standards. Companies invite testing teams; Genbounty triages; the company accepts; the team is paid in USDC from the company's own program balance or managed fiat
- **Reporting infrastructure for behavioural harm** - Purpose-built workflow from probabilistic evidence to remediation
- **Independent continuous testing** - External testers have no conflict of interest with the team that built the system
- **Two reporting paths** - Public reports on any AI product, plus scoped paid programmes, into one FLARE record
- **Two-sided marketplace** - Companies (buyers) and vetted AI safety testing teams (researchers)
- **Team-based model** - Companies invite teams, not individual researchers
- **Responsible disclosure** - Structured submission workflow with severity ratings and reproduction steps

## How It Works

1. **Scope** - The company defines the feature, risk categories, operating context, rules of engagement, and rewards
2. **Test** - Invited, vetted researchers run repeated trials under realistic conditions
3. **Evaluate** - Genbounty triages behaviour, reproducibility, impact, and safeguard failure
4. **Resolve** - The company documents, tests, accepts, and remediates. The locked USDC is released after the company and the team lead both confirm, and each tester collects their share

## Program Scope Types

- **Applications & Code** - AI-assisted codebases, web apps, APIs, and user-facing products (https://genbounty.com/ai-generated-applications)
- **Agents** - Autonomous agents with tool access and multi-step workflows (https://genbounty.com/bug-bounty-for-ai-agents)
- **RAG** - Retrieval-augmented generation pipelines and knowledge bases
- **MCP** - Model Context Protocol servers and integrations
- **Models** - Fine-tuned and hosted model endpoints
- **Prompts** - System prompts, guardrails, and prompt engineering surfaces

## Encrypted reporting

On invited company programs, exploit details are sealed on the researcher device before upload. The envelope is wrapped to Genbounty triage, the company Passkey, and the researcher Passkey. Title, severity, and program name stay readable so a finding can be routed. Opening sealed exploit details happens in the recipient browser.

- Companies: https://genbounty.com/support/get-started/encrypted-passkey-reports-for-companies
- AI whitehats: https://genbounty.com/support/get-started/encrypted-passkey-reports-for-ai-whitehats
- Public reports: https://genbounty.com/support/get-started/encrypted-public-ai-safety-reports

## Team Types

### External Teams
- Professional AI safety testing teams vetted and admin-approved by Genbounty
- Companies invite teams to specific private programs
- Teams build reputation through quality submissions and leaderboard ranking
- Browse teams: https://genbounty.com/teams

### Internal Teams
- Company-managed testers
- Pre-approved and controlled by the company
- Ideal for confidential or ongoing testing

## For Companies

- AI application safety and security testing: https://genbounty.com/for-companies
- AI Application Safety Platform: https://genbounty.com/ai-safety-bug-bounty-platform
- Browse AI safety testing teams: https://genbounty.com/teams
- White paper: https://genbounty.com/whitepapers/why-ai-safety-reporting-needs-its-own-bug-bounty-model

## Pricing

- **Points only (free):** Researchers earn reputation points for accepted findings, with no USDC bounty. Helps teams rank higher on https://genbounty.com/teams and qualify for paid projects.
- **Private pilot (250 USDC):** One deployed AI feature for 4 weeks.
- **Team hire:** Each team lists its own USDC rate per component, per team, per month. One unit is one team testing one AI component for one month.
- **LLM Hunter:** Individual lifetime license 5,999 USDC. Commercial lifetime license 59,999 USDC.
- Details: https://genbounty.com/pricing
- Sign up: https://genbounty.com/signup?role=company

## On-chain access model

Powered by Web 3.0. Safety testing by humans. The Genbounty platform is fully automated by agents and smart contracts, so human developers and testers can get to work. Genbounty has no access to the items below.

- **AI product testing** is private and invite-only. Genbounty does not publish the brief, the scope, or who you invited. Only your team and the testers you invite can open it.
- **AI safety reports** are encrypted in the researcher's browser before upload. Genbounty stores ciphertext. The key that opens a report lives in the company Passkey and is never written back to Genbounty servers.
- **AI security reports** seal reproduction steps, payloads, and evidence with AES-256-GCM on the tester's device. Genbounty keeps the ciphertext. Only a company Passkey can unwrap it.
- **Consumer reporting** encrypts sensitive answers and files in the browser before they leave the reporter's device. The sealed report is delivered to the company team. Genbounty does not hold the key.
- **Bug bounty payments** sit in the company's program balance contract. Genbounty never signs a transaction. The company and the testing team move the funds on-chain.

## Payments

Hired teams are paid in USDC on Base from the company's own program balance. Genbounty never signs a transaction and cannot move funds in or out.

- The company creates the program balance contract in a browser wallet on Base. That wallet is the owner. A little ETH on Base pays gas.
- Top up moves USDC from the owner wallet into the program balance.
- Hiring a team locks that team's listed USDC rate. One team, one AI component, one month.
- The USDC is released only after the company and the team lead both confirm the work.
- Each tester collects their share to a 0x USDC address saved in Security.
- Unlock returns a locked hire to the program balance when cancellation is allowed.

- Company program balance guide: https://genbounty.com/support/get-started/paying-a-hired-team-with-escrow
- Collect a hire payment: https://genbounty.com/support/get-started/collecting-hire-payment
- Program balance page: https://genbounty.com/company-escrow

## For AI Whitehats

- AI safety testing overview: https://genbounty.com/ai-safety-testing
- Sign up: https://genbounty.com/signup?role=ai-whitehat
- Browse teams to join: https://genbounty.com/teams/join
- LLM Hunter: https://genbounty.com/llm-hunter
- Academy: https://genbounty.com/academy

## Contact

- Sales: sales@genbounty.com
- Support: support@genbounty.com
- General: team@genbounty.com
- Investors: investors@genbounty.com
- Website: https://genbounty.com