Rosetta Daily · Sep 20, 2026
Seventy of today's 101 NVD entries are WordPress plugins, and three of them are the same mistake: Forminator, WP Recipe Maker and ProfilePress all run
do_shortcode()over input nothing validated. The highest-scoring one is not in that group — in Gravity Forms a hidden upload field walks past extension validation entirely.
Actively exploited
Orkes Conductor: unauthenticated remote code execution under active exploitation (CVE-2026-58138, CVSS v3.1 9.8 / v4 9.3)
Fortinet reports exploitation in the wild. Versions from 3.21.21 up to but not including 3.30.2 are affected, and an attacker needs no credentials to run code remotely. The CVE is not currently in the CISA KEV catalog.
Sources: The Hacker News
Critical vulnerabilities
Gravity Forms: hidden upload fields skip extension validation, giving unauthenticated arbitrary file upload (CVE-2026-84434, CVSS 9.8)
All versions through 3.1.0.4. The field validation pipeline and the file persistence pipeline disagree: hidden file upload fields bypass extension validation, and a rejected file's upload state is still passed intact to upload_file() with no re-check.
Sources: NVD
Botiga Pro: a REST route with no authorization check lets anyone set arbitrary WordPress options (CVE-2026-86591, CVSS 9.8)
Versions before 1.6.5. Unauthenticated users can write arbitrary values to arbitrary options, which leads to privilege escalation and full site takeover. The same route also accepts scripts that then run on every front-end page, and allows arbitrary posts to be moved.
Sources: NVD
Three plugins, one pattern: do_shortcode()run over unvalidated input (CVE-2026-92229, CVSS 9.1; CVE-2026-89274, CVSS 9.1; CVE-2026-85658, CVSS 8.1)
Forminator (through 1.57.2) lets unauthenticated attackers execute arbitrary shortcodes. In WP Recipe Maker (through 10.8.1), WPRM_Metadata::sanitize_metadata() recursively calls do_shortcode() on every scalar field of the recipe metadata, including reviewBody, which is filled verbatim from the text of approved rating comments. ProfilePress (through 4.17.2) is reachable from a subscriber-level account.
Welcomizer: the AJAX handler checks a nonce but not a capability, then hands user-supplied code to eval() (CVE-2026-4327, CVSS 8.8)
All versions through 2.8.1. The savesection handler in twiz-ajax.php verifies a nonce and never calls current_user_can(), while the contents of the "custom logic" field are executed with eval() on the front end.
Sources: NVD
UsersWP: social login never confirms who owns the email address, so anyone can log in as anyone (CVE-2026-86814, CVSS 8.1)
Versions before 1.5.10. The plugin resolves an existing account by email address without verifying that the social login provider confirmed ownership of it. An attacker who can assert an address through a provider account of their own logs in as that user, administrators included.
Sources: NVD
MgoSync: an unauthorized REST endpoint hands out the site's stored WooCommerce credentials (CVE-2026-92404, CVSS 7.5)
Versions before 2.1.7. Unauthenticated users can read the consumer key and secret, both of which carry read/write rights.
Sources: NVD
OpenShift console: /api/devfile/is reachable without authentication and yields SSRF (CVE-2026-75885, CVSS 9.3)
/api/devfile/ and /api/devfile/samples/ accept crafted devfile payloads from unauthenticated callers. The console pod then issues requests to internal services and reflects part of the response back. The same endpoints can also be exhausted by repeated large requests sent without a content length.
Sources: NVD
SolarWinds Access Rights Manager: a hard-coded key enables unauthenticated remote code execution (CVE-2026-28326, CVSS 8.8)
All versions of Access Rights Manager 2026.2 and prior are affected; SolarWinds has shipped updates.
Sources: The Hacker News
Totolink A3002MU: four remotely triggerable buffer overflows with public exploit code (CVE-2026-93741, CVE-2026-93740, both CVSS 9.3; CVE-2026-93739, CVE-2026-93738, both CVSS 8.6)
Firmware Hh-B20211125.1046. formWlWds, formWlEncrypt and formWlAc are reached through the submit-url argument, formSchedule through the webpage argument. Exploit code has been released publicly for all four.
SiYuan: notebook names and heading style attributes go unescaped, putting XSS inside the Electron renderer (CVE-2026-93923, CVE-2026-93922, both CVSS 8.6)
All versions through 3.8.4. The outline and bookmark docks fail to escape heading style attributes, and the Daily Note picker renders notebook names as raw HTML. The payload runs in an Electron renderer with Node.js access, which amounts to operating system command execution.
Acode: an exported Android service with no binding permission runs commands for any installed app (CVE-2026-68928, CVSS 8.6)
Versions 1.11.6 until 1.12.7. TerminalService is declared exported in plugin.xml without a binding permission and the service never verifies its caller, so any installed application can bind it and send MSG_EXEC with a command of its choosing.
Sources: NVD
Suricata: heap out-of-bounds read in the SMTP MIME quoted-printable decoder (CVE-2026-57228, CVSS 8.2)
Versions 7.0.13 until 7.0.17. When a quoted-printable escape sequence is split across traffic chunks and the following chunk holds exactly one byte, src/util-decode-mime.c reads one byte past a heap buffer. Crafted SMTP traffic triggers it — which puts the trigger in the hands of the party being inspected.
Sources: NVD
Microsoft Edge: use after free allows local privilege escalation (CVE-2026-88097, CVSS 8.1)
Sources: NVD
AI security
[Added detail] BragJack: one extension hijacks the built-in AI assistants of five browsers, two CVEs and over $20,000 in bounties
The attack comes from Gal Weizman at Forever Security and uses a technique he calls Prompt Forcing. It covers the assistants built into Chrome, Edge, Opera Neon, Perplexity Comet and Claude in Chrome.
Sources: Bleeping Computer
Gemini reached real corporate systems during a security evaluation
The incidents date to May 2026, in a test run conducted by the Israeli company Irregular, where a domain mix-up sent the model against companies that were never in scope. The same evaluation partner was involved in similar incidents disclosed earlier. First reported by The Wall Street Journal.
Sources: The Hacker News
Researchers used Claude Opus 5 to chain two flaws into OpenAI staff accounts and an internal repository
Three researchers at the security firm Hacktron chained a bug in the software running OpenAI's public help forum with a weakness in OpenAI's own login system, taking over the ChatGPT and Codex accounts of several OpenAI employees and reaching an internal code repository. This was security research.
Sources: The Hacker News
Supply chain
The second hop out of the TanStack poisoning: about 170 CrowdSec private repositories were copied four months ago
CrowdSec disclosed on September 18 that on May 22 an attacker copied roughly 170 of its private GitHub repositories using the account of an employee who had just left, whose access had never been revoked. That employee's laptop was compromised in May's TanStack npm supply chain attack, in which malicious versions of the packages stole credentials from developer machines.
Sources: The Hacker News