Bug bounty
Bug bounty roundup: 10 disclosed reports worth reading
Ten HackerOne reports from 2 Oct to 12 Nov 2026: an R3 Artifactory takeover, a MariaDB use-after-free RCE, a Fizzy token-stealing XSS and more.
This roundup collects ten bug bounty reports that HackerOne programs disclosed publicly between 2 October and 12 November 2026. Researchers found the bugs, the programs fixed them, and the write-ups became public, each credited to its named researcher and linked to the original report. The CVSS numbers are the Common Vulnerability Scoring System ratings the archives record. The archives held only a couple of write-ups from the last 48 hours, reports 9 and 10 here, so to reach ten worth reading we widened the window back to early October, skipping anything an earlier roundup covered.
Two themes run through the set. The larger is authentication and authorization that trusts the wrong thing: an Artifactory signing with an empty default key, a bbPress read path that forgot its own visibility rule, a MariaDB grant that outlives the object it protected, and a Fizzy page that mints a token on another account. The other is memory safety in MariaDB, where a cursor array frees a buffer still in use and a geometry function reads past its input, plus a backup extractor that writes outside its target. None of the ten lists a bounty.
Authentication bypass: a default join key hands over R3's Artifactory
qw3rjo reported that R3 Limited's JFrog Artifactory at software.r3.com ran with an empty join key, the shared secret the server uses to let internal services register. Under CVE-2026-82329 an empty key makes the signing secret 32 space characters, so the researcher forged a JSON Web Token (JWT, a signed token), signed it with an HMAC keyed hash from that known value, exchanged it for a service token, and minted an admin token. That reached 292 accounts, 269 repositories including Corda and central-bank-digital-currency builds, stored mail credentials and the GPG signing key, so an unauthenticated attacker could publish signed releases that banks pull. The fix sets a strong random join key and updates to Artifactory 7.191.14 or later. Program: R3. Severity: critical (CVSS 10.0). Bounty: not disclosed. Reporter: qw3rjo. Report: https://hackerone.com/reports/3994056
Access control: bbPress leaks private forum posts through the REST API
winstoncrooker found that the bbPress 2.7.0 development build exposed private and hidden forum content. bbPress stores topics and replies as public and marks only the parent forum restricted, enforcing visibility through a filter that runs on list queries. The 2.7.0 line newly published these post types on the WordPress REST API (representational state transfer, the HTTP data interface at /wp-json), whose single-item route reads a post by numeric ID with no query, so the filter never fires and the check passes anything public. An unauthenticated request by ID returned full secret content, and sequential IDs let the whole restricted set be walked. The released 2.6.x line is unaffected, so reporting it pre-release kept the regression out of production. The fix adds a REST permission callback that checks visibility. Program: WordPress. Severity: critical (CVSS 9.3). Bounty: not disclosed. Reporter: winstoncrooker. Report: https://hackerone.com/reports/3801107
Use after free: a MariaDB cursor array yields code execution
rdjgr chained two memory bugs into remote code execution (RCE, running attacker commands on the server) from an account with only per-database rights. MariaDB keeps open cursors inline in an array; opening the 33rd grows the array and moves its buffer while earlier cursors still point into the freed one, a use-after-free, so a later FETCH dispatches through a reclaimed object. A second bug, an out-of-bounds read in ST_Area when a MULTIPOLYGON declares more sub-polygons than it supplies, leaked memory to beat address randomisation. The researcher refilled the freed buffer with crafted bytes, flipped a privilege flag through the hijacked call, and ran a shell, needing no FILE or SUPER right. He scores it CVSS 8.5. Store cursor elements by pointer and bounds-check the geometry loop; fixing either breaks the chain. Program: MariaDB. Severity: critical. Bounty: not disclosed. Reporter: rdjgr. Report: https://hackerone.com/reports/3856148
Reflected XSS: one link steals a Fizzy user's write token
pirikara showed cross-site scripting (XSS, running attacker JavaScript inside a trusted origin) against Fizzy, Basecamp's board app, firing when a signed-in victim opens one crafted public-board link. The pagination helper passes every request parameter to Rails url_for, and the reserved script_name option rewrites the page-2 address to an Active Storage blob proxy. The attacker uploads an unattached HTML blob typed text/html;charset=utf-8, which slips past a binary-serving check that matches only bare text/html, and an unattached blob is authorized to any logged-in identity. Turbo fetches it as a page fragment and runs its script, copying the content security policy nonce (CSP, the rule set that would block injected script) onto it. The script mints a write token on the victim's identity and sends it to the attacker. The fix allowlists pagination parameters and binds each blob to its uploader. Program: Basecamp. Severity: high (CVSS 8.7). Bounty: not disclosed. Reporter: pirikara. Report: https://hackerone.com/reports/3943339
Broken access control: a dropped MariaDB package leaves its grant
heyoimjeb found that DROP PACKAGE, in MariaDB's Oracle mode, removes both routine rows but deletes only the PACKAGE row from the privilege table, leaving the PACKAGE BODY grant orphaned. The body gates execution, so a package of the same name created later runs for the old grantee under the new definer's privileges. The researcher granted a low-privilege user EXECUTE on a package body, dropped it, had root recreate a package of the same name reading a protected table, and watched that user call it and read a planted secret. The stale grant also flows through roles to current and future members. It needs an administrator to drop and recreate under a stable name, the normal deployment pattern. The fix deletes both privilege rows on drop. Program: MariaDB. Severity: high (CVSS 8.0). Bounty: not disclosed. Reporter: heyoimjeb. Report: https://hackerone.com/reports/3908943
Path traversal: a crafted backup archive writes files anywhere
riscj reported a Zip Slip class path traversal in mbstream, the xbstream extraction tool shipped with mariadb-backup. The extractor does not reject ".." components in an archive's chunk paths; its guard blocks only absolute ones. A chunk path of ../../etc/cron.d/evil expands under the target directory and the kernel resolves it to /etc/cron.d/evil, dropping the file outside the target. Restore jobs commonly run as root, and a new file in /etc/cron.d brings root command execution on the next minute. The researcher built a minimal archive with one traversing chunk and showed the cron job firing 40 seconds after extraction on Debian 12. These archives reach restore pipelines through backup buckets and automation the host does not fully trust. The fix anchors extraction to a fixed directory handle and rejects ".." before any syscall sees it. Program: MariaDB. Severity: high (CVSS 7.8). Bounty: not disclosed. Reporter: riscj. Report: https://hackerone.com/reports/3678395
Resource exhaustion: a Khan Academy page floods the visitor's browser
letx_world reported that a logged-in user who opens Khan Academy's /pwreset?action=reset endpoint sets the browser firing thousands of requests on its own, with no further interaction, until memory climbs and the tab crashes or is suspended within minutes. This is a client-side denial of service (DoS, making something unusable): the victim's browser is exhausted, not the server, so a single link in a phishing message could knock a target offline. The write-up is short on root cause, so treat it as a runaway request cycle on the reset flow and check the network panel for an endpoint that re-fires with no stop condition. The fix breaks the loop so a view cannot schedule unbounded requests. Program: Khan Academy. Severity: high (CVSS 7.7). Bounty: not disclosed. Reporter: letx_world. Report: https://hackerone.com/reports/3987499
Code injection: a malicious template runs commands through AWS Amplify
matty69v found remote code execution in amplify-velocity-template, Amazon's fork of velocityjs that the AWS Amplify command-line interface (CLI) uses to render VTL resolver templates (Velocity Template Language, the syntax AppSync resolvers use) during amplify mock api. The fork never backported upstream's prototype-chain guard, so property lookups resolve any inherited property, and an attacker-controlled .vtl file walks from an empty object through its constructors to the Function constructor to run operating-system commands. A poisoned resolver file arrives through a pull request, a developer runs the mock, and any GraphQL query spawns a child process with the developer's privileges, reading AWS credentials or SSH keys while the resolver still returns valid JSON. The package and the CLI that pulls it in see hundreds of thousands of weekly downloads. The fix backports the guard or allowlists safe property access. Program: AWS VDP. Severity: high. Bounty: not disclosed. Reporter: matty69v. Report: https://hackerone.com/reports/3897236
DOM-based XSS: a single Playground link runs script on wordpress.net
ariverapoblet reported a DOM-based cross-site scripting bug on playground.wordpress.net, where the sink sits in the browser's document object model (DOM, the in-page structure scripts read and write). The service worker intercepts any path ending /wp-includes/empty.html under a /scope:<id>/ prefix and returns a page that runs document.write on the decoded URL fragment, with no content security policy. The fragment after the "#" is attacker-controlled, so a crafted link writes arbitrary markup and script into the real wordpress.net origin. The payload lives in the fragment, so it never reaches the server and leaves no trace in logs or a web application firewall. Anyone who has opened Playground before has the service worker registered, so one click reads and exfiltrates the origin's cookies and local storage. A service worker that builds a page is as much an XSS sink as a server template. The fix stops writing the fragment unescaped and adds a policy. Program: WordPress. Severity: low. Bounty: not disclosed. Reporter: ariverapoblet. Report: https://hackerone.com/reports/3863343
Information disclosure: an edge cache serves one user's page to everyone
audion found that Fizzy marks its public board pages cacheable with Cache-Control: public for 30 seconds, but the controller that allows unauthenticated access still resumes the session, so a signed-in visitor's response carries their user id, cross-site request forgery token, CSP nonce and timezone. Nothing sets Vary: Cookie, so the content delivery network (CDN, the edge cache in front of the app) stores that personalised response under the plain URL and serves it to every later requester for the cache window. The proof started cold: one signed-in request, then two cookieless requests that came back byte-identical with the same user id and nonce. The CSRF token is bound to the session cookie, but the shared nonce stops being unpredictable and the user id reveals who viewed the board. The fix keeps personalised responses out of the public cache. Program: Basecamp. Severity: low (CVSS 3.1). Bounty: not disclosed. Reporter: audion. Report: https://hackerone.com/reports/4151790
What to take from this batch
Most of these are a check written once and not carried to every path that needs it. bbPress guarded its list queries but not the single-item read, MariaDB cleaned up grants on every drop except one, and Fizzy cached the right response for anonymous visitors but not for signed-in ones. When you fix access on one route, walk every other route that reaches the same object. The memory bugs point the other way, to a geometry loop and a cursor array trusting a length or a pointer they no longer own, which is what source review catches before it ships.
Our web application penetration testing works through signup, redirect, caching and authorization flows like the Fizzy, bbPress and R3 reports, and red team code review reads source for the incomplete fixes and missing bounds checks behind the MariaDB and Amplify bugs. Open the chat and Yaali, our AI agent, will pass your question to an engineer.
Sources: Individual reports on HackerOne: https://hackerone.com/reports/3994056, https://hackerone.com/reports/3801107, https://hackerone.com/reports/3856148, https://hackerone.com/reports/3943339, https://hackerone.com/reports/3908943, https://hackerone.com/reports/3678395, https://hackerone.com/reports/3987499, https://hackerone.com/reports/3897236, https://hackerone.com/reports/3863343, https://hackerone.com/reports/4151790. Archives used for ranking and detail: https://github.com/ajaysenr/HackerOne-Disclosed-Reports and https://github.com/Gh0stSpy/HackerOne-Disclosed-Reports-Database