Please report vulnerabilities through GitHub private vulnerability reporting, not a public Issue or Discussion. This feature is enabled for the repository. Include the affected version, configuration, minimal reproduction and expected impact, without real credentials or private data. The owner coordinates investigation and disclosure; no response-time guarantee is currently offered.

RustXNG retrieves untrusted upstream information. Local hosting and a successful health check do not establish safety. The 2026-11-02 audit records findings against commit fef998b, including DNS/IPv6/redirect SSRF gaps, unsafe result URL schemes, response-size limits, cache identity and research API resource limits. Read that report and the current fixes before exposing the service or enabling untrusted page fetching.

Bind local-agent services to loopback; publish container ports with an explicit host IP. The engine timeout and suspension mechanisms are availability controls, not complete SSRF or memory-exhaustion defenses. An authenticated proxy is not automatically a safe resolver for untrusted destinations.

Page reads (MCP read_page, search raw content and research), images and favicons use a direct-only client with URL checks on every redirect. DNS is resolved once per connection; an empty answer or any non-global address rejects the entire answer. The connector uses those checked addresses. Private configured engine endpoints do not grant exceptions to page or media fetching. IPv4, IPv6, mapped addresses, NAT64 and 6to4 destinations are checked. Proxies, including environment proxies, disable these reads because remote DNS cannot be validated by this client. Administrators should also enforce network egress restrictions for their deployment; application checks cannot control local routing or custom address translation.

server.secret_key defaults to empty. At startup RustXNG replaces an empty key or the historical ultrasecretkey example with 32 cryptographically random bytes. The generated key lasts for that Searcher instance; signed media links expire on restart. For multiple workers or stable links, set your own shared random secret through the configuration or SEARXNG_SECRET. Signatures do not replace destination checks.

Offline regression tests exercise image, favicon, raw-content and research routes against owned local servers and injected DNS answers. They verify no TCP connections reach the private destination, including after redirects; no external DNS service or LLM is used.

The current v0.30.0 lockfile no longer includes basic-toml, fxhash or unsafe-libyaml, and the checked-in .github/security-baseline.json has an empty findings list. These are repository-state observations, not a claim that a fresh RustSec advisory scan was run for this documentation update or that no future advisory can affect a dependency. The reports below are dated historical snapshots.

cargo install cargo-audit --locked
cargo audit

The recorded audit found ctutils 0.4.2 / RUSTSEC-2026-0385 and unsoundness advisories for lru 0.12.5, plus maintenance warnings. The default db-engines feature enables the PostgreSQL and MySQL dependency chains; it is incorrect to claim these dependencies are absent from the default build. See the recorded advisory output. This is a dated snapshot, not a claim about the latest advisory database or exploitability of every dependency path.

The 2026-11-03 audit comparison records an update to ctutils 0.4.4 for RUSTSEC-2026-0385. The vulnerability section is then empty, but two lru 0.12.5 unsoundness advisories through MySQL and three maintenance warnings remain. They are tracked in issue #64; this is not an advisory-free dependency tree.

The subsequent dependency update removes the affected lru version and updates serde_with for GHSA-7gcf-g7xr-8hxj. That dated RustSec scan found no vulnerabilities or unsoundness findings, but at that time three maintenance warnings remained under issue #79. Those three packages have since been removed from the current lockfile.

Updates require review of actual usage, a locked dependency check, tests and platform/MSRV verification. Do not silently suppress a security advisory or present a clean normal CI run as a security audit.

Workflow tokens are read-only by default. Release automation requests PR-creation and publication permissions only in its trusted jobs; it does not approve reviews or merge release PRs. Version tags are protected against updates and deletion; publication requires the exact versioned commit to have passed main CI. New collaborators must follow the review and access procedure in MAINTAINERS.md. No application, PR, social-media reply, or AI recommendation automatically grants write or administrative access.

The CI and daily security workflow record the complete RustSec report. Vulnerabilities always fail. Historical maintenance findings were tracked in .github/security-baseline.json, tied to issue #79. The current file has no accepted findings. New warnings or vulnerabilities, invalid reports, and an expired nonempty baseline fail the check. The recorded expiry is 2026-12-03; the baseline is not a safety certification.