SE Bug Bounty Arena / Report toolkit
Free report toolkit
AI bug bounty report toolkit
Everything I use to turn a finding into a report that gets accepted: a template to copy, a title builder, a severity guide for AI bugs, one finding written two ways, and a checklist to run before you submit.
- Free, no sign-up
- Built for AI and LLM bugs
- Copy and use
What a triager checks, in order
- Can I reproduce it?Steps
- Is it a real bug class?Title
- What does the attacker gain?Severity
- Is it written clearly?Example
- Is anything missing?Checklist
Each row jumps to the part of the toolkit that answers it.
The report template
Eight short parts, in the order a triager reads them. Copy it, replace everything in square brackets, and delete any part that does not apply.
Title [Bug class] in [feature or asset] lets [who] [do what] Bug class [For example: LLM01 Prompt Injection] Asset [The feature, endpoint or assistant where it happens] Summary [Two sentences: what is wrong, and why it matters] Steps to reproduce 1. [Where you start, and which account you use] 2. [The exact input you send, in a code block] 3. [What the application does] 4. [The response or result that proves the bug] Worked in [x] of [y] attempts. Impact [What an attacker gains, and who is affected. Name the data or the action.] Suggested fix [The change that removes the cause, not only this one input] Evidence [Screenshots or the copied response, with anything private removed]
Title builder
The title decides how fast your report is picked up. A good one says what the bug is, where it is, and what it lets someone do.
Severity guide for AI bugs
Rate what the attacker can reach, not how clever the prompt was. This is the scale I use as a starting point. A program's own severity table always wins.
| Severity | What it means | Typical AI examples |
|---|---|---|
| Critical | The AI feature can be made to run code or commands, or to take actions on other users' accounts at scale. | An agent with a code tool runs attacker-supplied code. An assistant can be steered into changing any customer's booking. |
| High | Reading or changing another user's data, or exposing a secret that gives real access. | A RAG assistant returns another tenant's documents. A leaked prompt contains a working API key. |
| Medium | A security control is bypassed with limited reach, or sensitive business logic is exposed. | A refund limit written only in the prompt is talked past. Model output is rendered as HTML for the attacker's own session only. |
| Low | Information of little value leaks, or a guardrail is bypassed with nothing further gained. | A system prompt with only tone rules is extracted. The assistant answers a question outside its topic. |
| Informational | Unexpected behaviour with no security effect. | The model gives a wrong or odd answer. A jailbreak produces text the attacker could have written themselves. |
Not sure? Answer three questions
Your suggested severity appears here.
One finding, written two ways
The same bug on a made-up travel assistant. The first version is what most first reports look like. The second is the one that gets paid.
Weak report
- Title
- Prompt injection!!
- Steps
- I uploaded a file and the bot did what the file said. Very dangerous.
- Impact
- Hackers can do anything.
- Fix
- Make the AI more secure.
Why it closes as "needs more information": no place, no input, no proof, and an impact nobody can check.
Strong report
- Title
- Indirect prompt injection in the itinerary summariser lets anyone who sends a traveller a PDF read that traveller's saved bookings
- Steps
- 1. Sign in as traveller A, who has two saved bookings. 2. Upload the attached PDF to "Summarise my itinerary". The PDF contains one line addressed to the assistant, pasted below. 3. The summary includes both bookings in full, formatted as the PDF asked. Worked in 4 of 5 attempts.
- Impact
- Anyone who can get a traveller to summarise a document can read that traveller's booking references, names and travel dates. No access to the account is needed.
- Fix
- Pass uploaded documents to the model as quoted data, and do not let document text trigger the bookings tool. Require the traveller to confirm before bookings are included in any output.
Why it resolves: a triager can repeat it in two minutes and can see exactly whose data is exposed.
Checklist before you submit
Nine checks. They are saved on this device, so you can come back to them.
Why reports close
The Arena uses the same four outcomes as real programs.
- Needs more informationThe triager could not reproduce it. Steps are missing, the input is described instead of pasted, or the starting point is unclear.
- Not applicableThe behaviour is real but is not a security issue for this program, or it is out of scope. A jailbreak with no further effect often ends here.
- DuplicateSomeone reported the same cause first. Different inputs that hit the same flaw count as one bug.
- ResolvedReproduced, in scope, with a clear impact. This is the only status that pays.
On real programs, test only what the program's scope page allows, and stop as soon as you have proof. Reading more data than you need to prove the bug can get a valid report rejected.
Questions people ask
What makes an AI bug report get accepted?
A triager must be able to reproduce the bug in a few minutes and see why it matters. That means exact steps, the real input you used, what happened, and what an attacker gains. Most rejected reports fail on one of those four.
Is a jailbreak a valid bug bounty finding?
Usually not by itself. Many programs treat getting a model to say something it should not as out of scope or low severity. It becomes a finding when it leads somewhere: another user's data, a secret that works, or an action the attacker should not be able to trigger.
How long should a bug bounty report be?
As short as it can be while still letting a stranger reproduce it. A clear title, one line naming the bug class, numbered steps, a short impact paragraph and a suggested fix is enough for most findings.
How do I rate the severity of an AI vulnerability?
Rate what the attacker can reach, not how clever the prompt was. Reading or changing other people's data and running actions or code rate highest. Leaking a system prompt with nothing sensitive in it rates low. Always check the program's own severity table, because programs differ.
Should I include the exact prompt I used?
Yes. Paste the exact input in a code block, plus the response that proves the bug. Language models do not always answer the same way twice, so say how many attempts it took and how often it worked.
Put it to work
Pick a target, find a bug with no hints, and write it up. Premium members get every report scored against the rubric with written feedback.