Skip to content

pgAdmin, URL, XML, and loopback-boundary batch

Source: GitHub Security Advisories, updated 2026-05-18: GHSA-4rhg-h8f2-v4jm, GHSA-hv9p-2pqf-r5w3, GHSA-hr4r-fwpv-c95j, GHSA-p58c-q354-6c4f, GHSA-6p2c-69cv-3fxq, GHSA-j74f-g7vx-fh4x, GHSA-hp84-p2gq-6fvr, GHSA-5rv5-xj5j-3484, GHSA-cr42-rg2m-mq4q, GHSA-5cvp-p7p4-mcx9, and GHSA-q2pj-8v84-9mh5.

This batch is durable because it clusters around admin tools and parsers that cross boundaries on behalf of trusted operators: session stores deserialize server-side data, file managers follow links, import/export helpers reach shells or SQL, URL clients mis-scope destinations, XML backends expand entities, and reverse proxies accidentally turn loopback-only trust into public API access.

What changed

  • pgAdmin 4 advisories covered FileBackedSessionManager deserialization, weak authentication throttling, symbolic-link path traversal in File Manager, LFI/SSRF, stored XSS in Browser Tree and Explain Visualizer, command injection in Import/Export query export, and SQL injection in the Maintenance Tool.
  • Faraday's prior fix for host scoping was incomplete for protocol-relative URI objects, so URL policy code could still treat attacker-controlled destinations as in-scope.
  • Docling's JATS XML backend was vulnerable to XML entity expansion/XXE-style attacks when parsing hostile journal XML.
  • Neotoma exposed Inspector/API access through a reverse-proxy loopback-auth bypass when deployments trusted loopback headers or routing assumptions too broadly.
  • Arcane added a reflected SVG color-parameter XSS path that can become admin account takeover when an administrator views the crafted route.

Operator triage

  1. Inventory exposed admin consoles and helper features:
  2. pgAdmin 4 deployments with File Manager, Import/Export, Maintenance Tool, saved sessions, or externally reachable login.
  3. Faraday consumers that enforce allowed_hosts, webhook allowlists, proxy scoping, or SSRF guardrails around user-supplied URI objects.
  4. Docling pipelines that ingest JATS/XML from uploads, feeds, papers, tickets, or untrusted storage.
  5. Neotoma/Arcane deployments behind reverse proxies, especially when admin users browse untrusted links.
  6. Treat pgAdmin findings as a single admin-tool boundary review, not isolated CVEs. A low-privileged database user who can reach file, export, or maintenance helpers may chain read, write, command, and SQL primitives.
  7. For pgAdmin session deserialization, rotate Flask/session secrets and invalidate existing server-side session files if untrusted users could influence session storage or after a suspected compromise.
  8. For Faraday and similar URL clients, test protocol-relative values (//host/path), mixed URI object/string flows, redirects, DNS rebinding, IPv4-mapped IPv6, link-local, loopback, and RFC1918 targets.
  9. For Docling, disable external entities and DTDs at the parser factory, then bound entity expansion, input size, nesting depth, and fetch behavior in a regression fixture.
  10. For loopback-auth designs, make the reverse proxy inject a signed internal identity header only after authentication, strip all inbound copies, and reject direct backend exposure.

Replayable validation boundaries

  • Admin file boundary: create symlinks, nested escapes, absolute paths, and encoded traversal paths in a disposable workspace; file managers must stay under the configured root after final path resolution.
  • Shell/SQL boundary: pass metacharacters and stacked-query probes through export and maintenance helpers; command execution must use argv arrays and database APIs must preserve parameterization.
  • URL boundary: normalize scheme-relative and object-wrapped URLs before policy checks; enforce policy on the final socket address after DNS and redirects.
  • XML boundary: include a harmless local entity and exponential-entity fixture; parsing must fail closed without outbound network fetches or excessive expansion.
  • Reverse-proxy boundary: send forged loopback, forwarded, and internal-auth headers from the public side; the backend must not infer trust from client-controlled headers.
  • Admin XSS boundary: render untrusted labels, colors, summaries, and SVG parameters as text or sanitize after all decoding and transformation steps.

Durable controls

  • Keep admin helper surfaces behind explicit role checks and separate high-risk file, import, export, and maintenance permissions from ordinary database access.
  • Never deserialize session or cache contents unless the storage path, format, signer, and object types are strictly controlled.
  • Resolve filesystem and URL targets to canonical final destinations before authorization or allowlist decisions.
  • Disable XML external entities by default and add parser-level resource caps for every document-ingestion backend.
  • Make reverse-proxy trust cryptographic or mTLS-backed, not source-IP or loopback-header based.
  • Add adversarial regression tests for every bug class fixed here; most of these were bypasses of controls that looked present but were applied at the wrong boundary.