I started bug bounty hunting with no security background. Before I touched a target, I read public writeups, the reports researchers publish after a program pays them. After a while the same patterns kept showing up. These are the twenty that stuck with me.
1. A role without a scope
One writeup described a GraphQL mutation that checked the JWT for a supplier admin role but never matched it to the supplier ID in the body. An admin on supplier A became an admin on supplier B. My test: two accounts, two tenants, one privileged call replayed across them.
2. Hunt the legacy switch
A patch does not kill the chain. In one writeup, 19 of 21 servers were already on the hotfix, but the exploit chain needed the Legacy API switch turned on, and that switch was the whole vulnerability. On every target I check what old feature is still on.
3. Verification discipline
A launched process, a 403, a token receipt. None of these is proof. A candidate is confirmed only when a negative control fails and the positive replay succeeds twice. Many stacks put a parser in front of auth, so a malformed body gets a 400 first. Re-test with a well-formed body.
4. Write reports for a tired volunteer
One program nearly died because about 20 percent of its reports were AI slop and the confirmed rate fell below 5 percent. Every report needs five things: a clear title, reproduction steps, a proof of concept, an impact statement, a suggested fix. Before I submit anything I ask whether a tired volunteer would be glad to receive it.
5. Out-of-scope bugs escalate
The bug class is not banned. What gets rejected is the finding reported on its own. A missing rate limit becomes a finding when you show the math: a 6-digit code, a 10-minute window, no lockout. A CORS misconfiguration becomes one only when an endpoint returns authenticated data cross-origin with credentials included.
6. Model output is untrusted
Model output is untrusted. It must never grant authority, bypass policy, or trigger a privileged action without independent application controls. The question is whether untrusted content can change what the application is allowed to do. The finding is a missing control on that chain.
7. A jailbreak alone does not pay
A jailbreak by itself is not a finding. Most programs list jailbreaks out of scope. The paid prompt-injection reports lead somewhere: another user's data leaked, an unauthorized action happened, or the injection reached outside the chat. Write the report around the missing control, not the prompt.
8. Mass assignment
The cheapest critical I read about was one extra key. A PUT endpoint bound the whole JSON body to the database model with no allow-list, so adding "role": "admin" made a standard user an administrator, CVSS 9.9. On every PUT or PATCH endpoint I add privilege fields to the body: role, is_admin, is_staff, permissions, tenant_id, plan, balance. Then I prove it with a write in the admin surface.
9. Read the whole API response
One finding started with an endpoint that returned full emails and internal IDs for everyone enrolled, including people who never posted. About 14,000 emails came back in one response. Registration had no email verification, so any fresh account could harvest the whole list. I read every API response in full and ask whether the data is scoped to the requester.
10. Pre-auth flow steps are an API surface
One login flow ran its duplicate-check step before any authentication and returned the matching account's numeric user ID for any email or phone. My test: submit an identifier for a normal account, a hidden account, a suspended account, and a nonexistent one, and diff the responses. A pre-auth step must return a decision, never an identity.
11. Predictable randomness is not a secret
One writeup showed Math.random() used as an authentication secret, chained with a DOM XSS in a postMessage handler, ending in account takeover. I grep every JavaScript bundle for Math.random(), rand(), and random.random used to build tokens and secrets. If it is predictable, it is not a secret.
12. OAuth redirect chains
One login endpoint was a valid OAuth redirect URI. After a login migration, it bounced through the new login and back to a caller-supplied next URL. The old version did the final hop in JavaScript. The new version redirected directly, and fragment tokens travel across redirects. I map every redirect_uri and next parameter, and follow the chain with a fragment token.
13. Proxied user input
A proxy took an item ID from the URL and dropped it into the internal path. The WAF blocked the obvious ../, and a backslash variant slipped past. Error shapes did the mapping: 'not found' meant the internal API was reached. From there the researcher reached a search API over about 100 million customer records. On any proxy endpoint I try ../, ..\, %2e%2e, %5c, and ..;/, and diff the error shapes.
14. Internal-sounding surfaces are leads
Companies forget things: old test pages, unused login panels, internal portals on public subdomains. In one writeup, a portal's API returned full account data when a phone number was appended to the URL, with no auth. Any subdomain or endpoint that sounds internal is a lead. I hit it unauthenticated first.
15. Hunt the forgotten asset
The main app has had a hundred hunters on it. The old staging site has had none. First I read the scope and confirm the asset is in scope. Then I look at the forgotten surface before the homepage: staging subdomains, test pages, legacy panels, decommissioned tools, archived docs.
16. Spot IDOR without a second account
With only one account, generate the other user's identifier from data the app already shows: reviews, comments, listings, and avatar URLs. In one demo, a review exposed another user's avatar URL, whose path contained his user ID, and opening it leaked his information. A UUID or non-integer ID in a URL is not proof of safety.
17. One file for self-hosted reports
Without a platform, do not paste the full report into the email body. Send one file instead, a PDF or a .txt, attached to a short email. A self-hosted team triages from an inbox, and one clean file reads better than a long email body.
18. No bounty program? Get permission first
With no formal bounty program, do not just start testing. Find the official contact and ask for permission to test first. Only then ask about the reward or recognition. Permission comes first, so the testing is authorized. The reward question comes second, so it does not read as extortion.
19. Headers are attacker input
This one has a name and a number: CVE-2026-18516, Crafty Controller, CVSS 9.6, fixed in 4.11.0. Crafty stored the 'IP' from X-Forwarded-For and X-Real-IP in its logs without validating them. A failed login with a JavaScript payload returned a clean 401, but the payload was already saved. The log page inserted it into HTML unescaped, so it ran in the admin's browser when opened. My test: a harmless marker in those headers, a failed login, then open the log as an admin.
20. XInclude is XXE without a DOCTYPE
A 'Check stock' feature embeds productId in server-side XML. No DTD can be defined, because the attacker does not control the whole document. But the parser still processes XInclude directives, so an injected xi:include retrieves the file anyway. The payload is:
<foo xmlns:xi="http://www.w3.org/2001/XInclude"><xi:include parse="text" href="file:///etc/passwd"/></foo>
The detail that matters is parse="text": without it, /etc/passwd fails as invalid XML. Blocking DOCTYPE does not close XXE.
None of these twenty came from a course. They came from other people's writeups. Reading is the cheapest part of this work. The expensive part is testing each pattern with two accounts and a proof of concept.
Top comments (0)