SE Bug Bounty Arena / Real programs
Real programs, official links
AI bug bounty programs you can hunt on
Real programs that pay for vulnerabilities in AI products, and what each one says about prompt attacks. The Arena is practice. These are where the skill gets paid.
- 7 programs
- Each entry says where its details came from
How they treat prompt attacks
- Accepts prompt attacks2 programs
- Only with real impact2 programs
- Not prompt attacks2 programs
- Depends on the track1 program
Most programs pay for what a prompt attack leads to, not for the attack itself.
SecurityElites is not affiliated with any of these programs. Scope and rewards change often, so read the official page before you test anything. Each entry says where its details came from.
- Not prompt attacks
Google AI Vulnerability Reward Program
Google, on Google Bug HuntersWhat it covers. Security bugs in AI features of Google Search, Gemini apps and Workspace, with other Google AI products in lower tiers. Launched in October 2025.
What to know. Direct prompt injection, jailbreaks and alignment issues are excluded. Content problems are reported inside the product instead.
Rewards: Up to $30,000
Details from: Launch coverage from October 2025.
Open the official page - Only with real impact
Microsoft Copilot Bounty Program
Microsoft, on Microsoft Security Response CenterWhat it covers. Copilot on the web, in Edge, in Windows, in the mobile apps and on WhatsApp and Telegram. Covers classic web flaws plus inference manipulation and inferential information disclosure.
What to know. Prompt injection that affects nobody but the attacker, attempts to leak the system prompt, and model hallucination are excluded.
Rewards: $100 to $30,000
Details from: The official program page, read on 16 November 2026.
Open the official page - Only with real impact
OpenAI Safety Bug Bounty
OpenAI, on OpenAIWhat it covers. Announced in March 2026 for abuse and safety issues that are not classic security flaws. Covers third-party prompt injection that reliably hijacks an agent into a harmful action or a data leak, exposure of proprietary information, and ways around account and platform controls.
What to know. Jailbreaks are out of scope, and an agent hijack must reproduce at least half the time. Access to data or features beyond your permissions goes to the separate security bug bounty on Bugcrowd. The announcement gives no reward amounts.
Details from: The OpenAI announcement, read on 16 November 2026.
Open the official page - Accepts prompt attacks
Anthropic Model Safety Bug Bounty
Anthropic, on HackerOneWhat it covers. Universal jailbreaks that get past the Constitutional Classifiers on deployed systems. Entry is by application and invitation, and participants sign a non-disclosure agreement.
What to know. Infrastructure flaws such as cross-site scripting or SQL injection are out of scope here and go through the separate responsible disclosure policy.
Rewards: Up to $35,000
Details from: The official help page, updated 8 September 2026, read on 16 November 2026.
Open the official page - Accepts prompt attacks
0DIN GenAI Bug Bounty
Mozilla, on 0DINWhat it covers. Flaws in widely used generative AI models and applications from several vendors. Model bugs include prompt extraction, guardrail jailbreaks and interpreter jailbreaks. Application bugs must start with a prompt injection and lead to data being read, changed or code being run.
What to know. The bug must affect the latest generation of the model. Bounties are set at the discretion of the 0DIN team.
Rewards: $500 to $15,000
Details from: The official policy page, read on 16 November 2026.
Open the official page - Not prompt attacks
huntr
Palo Alto Networks, on huntrWhat it covers. Flaws in AI model file formats, such as unsafe deserialisation and backdoors. Pays up to $1,500 for pickle files and up to $3,000 for other formats. The platform is also moving to challenge-based competitions on AI systems.
What to know. The open-source software program stopped taking new reports on 30 June 2026. The work here is file-format flaws, not chat attacks.
Rewards: Up to $3,000 for model file formats
Details from: The huntr model format page and its 2.0 FAQ, read on 16 November 2026.
Open the official page - Depends on the track
xAI
xAI, on HackerOneWhat it covers. xAI names this HackerOne program as its bug bounty in its security.txt file.
What to know. We confirmed the program exists but could not read its scope page, so check there for which Grok assets and bug types count and what they pay.
Details from: The xAI security.txt file, read on 16 November 2026.
Open the official page
No program in this group.
Before you report to a real program
- Read the scope firstMost AI programs pay for what a prompt attack leads to, not for the attack itself. Know which kind you are hunting on.
- Check the rules on toolsSome programs limit automated testing. An AI assistant acting for you counts as a tool.
- Stop at proofTake only the evidence you need. Do not read, keep or share other people's data.
- Write it properlyA reproducible report with a clear impact is what gets paid. Use the toolkit before you submit.
Get ready first
Learn to read a program in five minutes, then practise the whole loop on a blind target.