Skip to content

Soul-driven AI agent with permission-hardened tools, token budgets, and multi-channel access. Runs 24/7 from CLI, Telegram or More.

1.6 KB
Raw Blame

Response to SSRF advisory (issue #108)

Status: Fixed — current releases not affected. Issue closed as fixed.

Summary

A responsible-disclosure report (issue #108, dated 2026-08-06) described authenticated SSRF via the fetch_url tool: model-driven fetch() with no destination filtering, allowing a chat user to probe loopback / internal endpoints and reflect response bodies back into the conversation.

  • Affected: ≤ 1.1.13 — fetch_url called plain fetch(url) directly.
  • Fixed: tool routed through the SSRF guard (src/utils/ssrf.ts) in the 1.2.3 release; hardened via #121 and #124.

Guard coverage

Vector Handling
Loopback / RFC 1918 / link-local (incl. 169.254.169.254) / CGNAT / multicast blocked
localhost, .local, .internal hostnames blocked
Non-http(s) schemes blocked
DNS answers containing any private record fail-closed
Redirects every hop re-validated (manual redirect mode)
DNS rebinding validated address pinned to the socket

Verification

The report's PoC was reproduced against main: a loopback canary was never contacted; all four attempted vectors were rejected by the guard. A sweep of remaining fetch() call sites found only one raw fetch (Slack file upload), which targets a URL issued by Slack's API — not model-controllable input.

Maintenance notes

  • guardedFetch is the only sanctioned path for model-reachable fetches.
  • Re-run a sink sweep (rg "fetch\(" src --type ts) when adding new network surfaces; local-provider base URLs are exempt by design (loopback trust boundary, web server binds 127.0.0.1 only).