CVE-2026-130358: AdminNeo's Claude, Gemini and Open WebUI SQL plugins ran on the login page, giving anyone an open relay to the operator's AI key (fixed 5.9.1)
AdminNeo, a PHP database manager, has optional AI SQL-helper plugins (ClaudeSqlPlugin, GeminiSqlPlugin, OpenWebUiPlugin). Before 5.9.1 they did their work in sendHeaders(), which also runs when the login page is rendered, and never checked for a session or CSRF token. An unauthenticated visitor could POST a claude, gemini or openwebui field and get the model's answer back, using the operator's Anthropic or Gemini API key or Open WebUI bearer token. For Open WebUI that also meant reaching a possibly internal endpoint with the operator's token. Fixed in 5.9.1 (16 November 2026).
What it is
page_headers() in admin/include/design.inc.php calls each plugin's sendHeaders() on every page, including print_login_page(); the plugins forward the POSTed text to api.anthropic.com (x-api-key), Gemini, or the configured Open WebUI /api/v1/chat/completions (Authorization: Bearer) and echo the reply.
The reporter (AlpetGexha) verified it end to end with no cookie or token: POST body openwebui=list all users returned the model's reply and the endpoint received the operator's bearer token.
Impact: arbitrary prompts billed to the operator, model and configuration disclosure, and an SSRF-like path to internal Open WebUI. The plugins are off unless the operator enables them.
How to check
- If you enabled any of the three AI plugins on AdminNeo below 5.9.1, you're affected. Check the AI provider's usage logs for requests you don't recognise.
How to fix
- Upgrade AdminNeo to 5.9.1 or later.
- Rotate the Anthropic/Gemini API keys or Open WebUI token configured in the plugins if the AdminNeo login page was reachable by untrusted users.
- Keep database admin tools behind VPN or an authenticating proxy.
composer show adminneo-org/adminneo 2>/dev/null | grep -i versions
grep -rlE "ClaudeSqlPlugin|GeminiSqlPlugin|OpenWebUiPlugin" . 2>/dev/null Sources
- criticalCVE-2026-63377 and CVE-2026-63378: TAKO AI Agent for Okta lets anyone grab the first admin account and any user run Python as root, upgrade to 3.0.1-beta
- criticalMalicious PyPI package aanthropic typosquats Anthropic's Python SDK and runs malware on install
- criticalMalicious npm packages posing as AI agent tools: agentic-awesome-skills, x402-prompt-injection-scan and formatting-subtitles-for-claude
- criticalLiteLLM 1.107.0 on PyPI and Docker was malware: a credential stealer that runs whenever Python starts