Supply Chain Guard

Version updated for https://github.com/homeofe/supply-chain-guard to version v5.12.0.

  • This action is used across all versions by 0 repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

What’s Changed

v5.12.0 (2026-07-11)

Issue #54 hardening: oversized-file transparency + threat-intel indicator contract

Implements both hardening gaps tracked in issue #54 (follow-up to the merged PR #55 extraction/IOC hardening), plus the dependency maintenance merged this cycle (docker/login-action 4.4.0, vitest + @vitest/coverage-v8 4.1.10). This minor also carries PR #55’s archive-extraction and self-scan-suppression fix to npm (it landed after v5.11.1 was published).

  • New rule FILE_TOO_LARGE_SKIPPED (info): the core, VSIX, npm, and PyPI scanners no longer skip files above the 5 MB content-scan limit silently - every skipped scannable file is surfaced with its path and size, because an attacker can deliberately pad a payload past the limit to dodge scanning. info severity: never affects exit codes; filterable via –min-severity or –exclude FILE_TOO_LARGE_SKIPPED. The oversized body is never read.
  • Threat-intel indicator contract: feed values are LITERAL indicators, never regexes. Domain values were previously compiled to RegExp with only dots escaped, so a hostile or malformed remote feed value like “(” threw inside the per-file loop - swallowed by the per-file catch, silently disabling all downstream checks for every file while the scan exited green, and a valid catastrophic pattern ("(a+)+b") would be ReDoS-tested against full file contents. Now: full metacharacter escaping (values match only themselves), compiled once per unique value, with a substring fallback that can never throw.
  • Type-aware quarantine at every feed ingestion point (feed refresh, the legacy update API, and the cached-feed load at scan time): each entry must match its type’s shape (domain/ip/url/hash/package charsets, 2048-char cap). Invalid entries are dropped deterministically; a rejected refresh never overwrites the previous cache. This also stops a structurally-valid garbage literal like a bare “(” from flooding reports with false matches.
  • npm/PyPI extracted-file walkers are now exported (scanExtractedNpmFiles / scanExtractedFiles) so the size-limit behavior is regression-tested without network; 16 new tests across all five scanner families and all three ingestion paths.
  • An adversarial review gate BLOCKED the first candidate with 6 confirmed findings, all fixed pre-tag: (1)+(2) the ip/url value shapes were charset-only, so a degenerate flood value like ip “.” or url “(” passed the quarantine and substring-matched a critical finding onto virtually every scanned file - ip now requires IPv4/IPv6 structure, url an 8-char floor; (3) the domain regex cache was unbounded (long-running MCP server + rotating hostile feed = monotonic memory growth) - now cleared at 10k entries; (4) severity is enum-checked and confidence range-checked (an unknown severity string would have produced NaN scores downstream); (5) the skills-scanner’s agent-rules reader also surfaces oversized files now (fifth family, full DoD parity); (6) the feed-reject error message bounds the attacker-controlled type field, not just the value.