The security docs claimed the in-browser SSRF filter (context.route/
routeWebSocket) re-applies "the same filtering" as the pre-fetch DNS
check. That's inaccurate for hostnames: the browser-level filter only
blocks literal private IPs and localhost/.local/.internal suffixes —
it never re-resolves hostnames, so a same-hostname DNS-rebind (public
IP on first resolve, private IP on a later request from inside
browserless) is not closed at that layer. Correct the wording in the
design spec and TECHNICAL.md, and add a prominent note to both the
spec's deploy section and the plan's deploy notes: before flipping
TOOLBOX_VISIBLE on prod, harden the browserless container's network
egress (block 169.254.0.0/16 and RFC1918 ranges via host firewall or
a dedicated internal docker network) to close the residual at the
network layer. Also note that per-user limits currently count only
successful generations — failed renders are uncapped, a bounded
self-DoS risk worth a follow-up.
Staging runs at staging.school.second-brain.ru on Hetzner port 3011.
Deploy script: ~/Documents/Claude/scripts/deploy-staging.sh
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
admins