CVE WatchAug 24, 2026

CVE-2026-78205

BentoML's outbound connection safeguard (make_safe_connect in _internal/utils/uri.py) blocks private, loopback, and link-local IP addresses

CVSS

5.8

Medium

EPSS

—

KEV

—

Exploit Today

0-100

Published: Aug 24, 2026 · Last modified: Aug 24, 2026 · CWE-918

EPSS · 30d

Not enough EPSS history yet.

Technical description

BentoML's outbound connection safeguard (make_safe_connect in _internal/utils/uri.py) blocks private, loopback, and link-local IP addresses but fails to reject the RFC 6598 shared address space (100.64.0.0/10, CGNAT). In versions 1.4.19 through 1.4.39, an unauthenticated attacker can supply URLs pointing to that range via multipart file handling (MultipartSerde.ensure_file) or JSON request parsing (JSONSerde.parse_request), causing the server to make outbound requests to internal hosts on CGNAT networks (Server-Side Request Forgery). This is an incomplete fix for CVE-2025-54381.

Official references

Related CVEs

CVECVSSEPSSKEVRExploitTitleMod.

CVE-2026-713667.7 HIG

—

———A server-side request forgery (SSRF) vulnerability was found in multiple AWX notification backends. The webhook, Mattermost, Rocket.Chat, and Grafana notification backends use notification template URLs as direct HTTP request targets without validating the target address against private, loopback, or reserved IP ranges. An organization notification administrator can create notification templates pointing to internal or loopback addresses, causing the AWX control node to issue HTTP requests to services that are not externally accessible. Additionally, the webhook notification backend follows HTTP redirects and resends configured Basic Authentication credentials to redirect targets regardless of host change, allowing an attacker to exfiltrate notification credentials by redirecting to an attacker-controlled host. The Grafana backend sends its API key in the Authorization header to the configured target URL.3h CVE-2026-78385—

—

———RansomLook contains insufficient resource validation in the analysis PDF generation functionality. Analysis documents are converted from Markdown to HTML and passed to WeasyPrint for PDF rendering. Prior to the fix, WeasyPrint used its default URL fetcher, allowing resource references contained in an analysis to be resolved without restrictions. An authenticated attacker able to create or modify an analysis could embed crafted resource references using schemes such as file:// or http://. When the analysis was subsequently rendered as PDF, WeasyPrint would process these references with the privileges and network access of the RansomLook server. A malicious file:// reference could cause the renderer to access arbitrary files readable by the RansomLook process, potentially exposing sensitive configuration, credentials, or other local data through rendered resources. Network URLs could cause the server to initiate requests to localhost, internal network services, or external systems, resulting in server-side request forgery (SSRF) and potentially bypassing network-level access restrictions. The patch introduces a dedicated WeasyPrint URL fetcher that permits only data: resources, the RansomLook report logo, and files contained within the analysis asset directory. Network resources and filesystem paths outside these explicitly permitted locations are rejected.4h CVE-2026-782774.9 MED

—

———Subscriber Server Side Request Forgery (SSRF) in FluentCRM Pro <= 3.1.12 versions.2h CVE-2026-782696.4 MED

—

———Contributor Server Side Request Forgery (SSRF) in Shared Files <= 1.7.69 versions.2h CVE-2026-105827.4 HIG

—

——0Hugo's security.http.urls allowlist is the only control on outbound fetches made by resources.GetRemote, and it inspects the URL text alone. CheckAllowedHTTPURL in config/security/securityConfig.go applies the configured pattern list and then re-checks a canonicalised form of an integer, hex or octal IPv4 host, but it never resolves the hostname and never inspects the address the HTTP client actually connects to. The client constructed in resources/resource_factories/create/create.go installs no dial-time hook, so no check occurs at connection time either. A hostname that resolves to a loopback, private or cloud-metadata address therefore satisfies the policy, and the response body is embedded in the generated site. An attacker who can supply a URL through content, for example a front-matter field or a CMS field, can make the build fetch an internal endpoint and publish the response in the static output, so the build artifact itself carries the data out.8h CVE-2026-759757.5 HIG

—

——0fast-uri is a URI parser for Node.js. Its custom parser for bracketed IPv6 literals does not validate the complete IPv6 grammar, so invalid trailing text in an authority can be silently discarded and a malformed attacker-controlled host is turned into a different valid IPv6 destination. For example, a bracketed literal with invalid trailing characters is normalized to the unspecified address, which a Node HTTP client then connects to a local service over loopback, and other malformed literals collapse to private-range addresses. No error is set on the parsed result, so an application checking the error field cannot detect the rewrite. An application that normalizes untrusted URLs before outbound requests, redirects, proxy routing, or address-policy enforcement can be redirected to a local or private IPv6 target, giving a server-side request forgery and address-policy bypass primitive. The affected versions are 2.3.1 up to but not including 2.4.5, 3.0.0 up to but not including 3.1.6, and 4.0.0 up to but not including 4.1.3. The issue is fixed in 2.4.5, 3.1.6, and 4.1.3, which validate bracketed IP literals against the full grammar and mark malformed literals as authority errors. Users should upgrade to a patched version.4h