Release v5.15.2 - security fixes · Jovancoding/Network-AI

Security fixes

GHSA-9p2w-prp8-5722 (High) - ClaudeHookBridge deny patterns now inspect every string field of tool_input (e.g. Write.content, Edit.new_string, MCP tool arguments), not just the first candidate field. Over-nested or oversized inputs fail closed. Reported by zx (Jace) / @manus-use.
GHSA-hr6v-mfxm-4438 (Moderate) - DashboardServer now validates the Host header (blocks DNS rebinding) and the WebSocket Origin (blocks cross-site WebSocket hijacking). New allowedHosts / allowedOrigins options for proxied deployments. Reported by zx (Jace) / @manus-use.
GHSA-4pvg-m42h-c3x2 (Low) - McpSseServer reflects a localhost CORS Origin only when bound to a loopback address. Reported via Sebastion AI (@andesyteoss).
CodeQL #180 (js/regex-injection) - hook deny/allow regexes are validated at construction (invalid, oversized, or nested-quantifier patterns rejected); network-ai hook pre-tool-use exits 2 on any error so a bad rule blocks instead of silently allowing.
CodeQL #179 (js/incomplete-sanitization) - scripts/clawhub-publish.js rejects arguments cmd.exe cannot safely quote and no longer uses a shell on non-Windows platforms.

Deny patterns now also see file contents and tool arguments, so a rule like --deny "rm -rf" will block writing a file that contains rm -rf.
A DashboardServer behind a reverse proxy with a different hostname must set allowedHosts / allowedOrigins.

OpenSSF Scorecard: ossf/scorecard-action v2.4.4 (image moved from gcr.io to ghcr.io).
Examples load dotenv via its typed main entry (dotenv 18 compatibility).

3,673 tests across 41 suites.

Full changelog: https://github.com/Jovancoding/Network-AI/blob/main/CHANGELOG.md