Skip to content

CHANGELOG.md

Latest commit

 

History

History
193 lines (91 loc) · 10.2 KB

CHANGELOG.md

File metadata and controls

193 lines (91 loc) · 10.2 KB

Changelog

0.6.3 (2026-04-23)

Bug Fixes

  • ENG-12707, ENG-12708: make boundary annotation opt-in (annotateBoundary flag) (#57) (bf10849)

0.6.2 (2026-04-21)

Bug Fixes

  • ENG-12702: strip boundary markers from input before classification (#55) (0fdd9d4)

0.6.1 (2026-04-21)

Features

  • ** ENG-12684:** v4 ONNX model + cumulative risk fixes + opt-in SFE preprocessor (includes ENG-12671) (#52) (6fb1ff9)

Miscellaneous Chores

0.6.0 (2026-04-16)

Features

  • ENG-12658: sentence density adjustment to reduce email notification FPs (#50) (b4a272d)

Bug Fixes

  • ENG-12643: prevent prototype pollution during object traversal (#47) (97f85b7)

0.5.8 (2026-04-08)

Bug Fixes

  • ENG-12604: chunk batch classification to bound ONNX memory usage (#44) (46e6548)

0.5.7 (2026-04-08)

Bug Fixes

  • ENG-12594: remove per-tool sanitization rules (#42) (eb4664b)

0.5.6 (2026-04-01)

Bug Fixes

  • ENG-12540: prevent unhandled rejection cascade on ONNX load failure (#37) (52c99e3)

0.5.5 (2026-03-31)

Bug Fixes

  • ENG-12518: fix field filter and batch sentence classification (#35) (878b542)

0.5.4 (2026-03-30)

Bug Fixes

  • ENG-12506: cache ONNX session at module level to eliminate per-request (#33) (bd1c580)

0.5.3 (2026-03-27)

Bug Fixes

  • ENG-12470: update inline docs to match implementation, scope Tier 2 to Tier 1 risky fields by default (#30) (c103072)

0.5.2 (2026-03-26)

Bug Fixes

  • use absolute URLs for README images to fix npm display (#28) (beeb2ef)

0.5.1 (2026-03-25)

Bug Fixes

  • ENG-12448: upgrade ML classifier to jbv2 (AgentShield 73.7 → 79.8) (#25) (3061239)

0.5.0 (2026-03-23)

Features

  • ENG-12396: upgrade ML classifier to jbv5 (AgentShield 73.7 → 81.1) (#22) (51f50ce)
  • ENG-12397: remove legacy MLP inference mode (#23) (556dc38)

0.4.5 (2026-03-12)

Bug Fixes

  • downgrade nanoid to 3.3.11 to fix CJS compatibility (#19) (06c1713)

0.4.4 (2026-03-11)

Bug Fixes

  • ENG-12332: export ToolSanitizationRule and add tier2Fields option (#16) (90bfa68)

0.4.3 (2026-03-09)

Bug Fixes

  • ENG-12109: sync toolRules passthrough fix and test updates from connect (#13) (f1eb775)

0.4.2 (2026-03-03)

Bug Fixes

  • DEF-1: document useDefaultToolRules and add ATS/CRM tool rules (53764e6)
  • ENG-12236: document useDefaultToolRules and add ATS/CRM tool rules (a563ad3)

0.4.1 (2026-03-03)

Bug Fixes

  • DEF-1: fix tsconfig extending missing monorepo base and add biome linter config (44993ba)
  • DEF-1: fix tsconfig extending missing monorepo base and add biome linter config (2469987)

0.4.0 (2026-03-03)

Features

  • add biome configuration file and fix code formatting (219526f)
  • initial release of @stackone/injection-guard v0.1.0 (039db9e)
  • v0.2.0 — ONNX-based Tier 2 classifier, API improvements (bbb204d)

0.3.1 (2026-02-19)

Bug Fixes

  • ENG-12119: release v0.3.1 - add SSPL-1.0 LICENSE and update docs (#767) (d250563)

0.3.0 (2026-02-19)

Features

  • ENG-12119: add @stackone/defender package (#747) (55f4ffb)

Bug Fixes

0.2.0

Added

  • ONNX-based Tier 2 classifier. Fine-tuned MiniLM-L6-v2 model exported to ONNX with int8 quantization (~22MB), bundled in the package. Now the default Tier 2 mode (mode: 'onnx'). Significantly more accurate than the previous MLP approach — 2-bench avg F1 0.876 vs 0.70.
  • defendToolResults() batch method. Defends multiple tool results concurrently via Promise.all.
  • fieldsSanitized and patternsByField in DefenseResult. Provides per-field observability into which fields triggered sanitization and which patterns were found in each.
  • Tier 2 lazy loading. ONNX model auto-loads on first defendToolResult() call if warmupTier2() wasn't called explicitly. warmupTier2() is still recommended to avoid first-call latency.

Changed

  • defendToolResult() is now the primary API. Single method that runs Tier 1 pattern detection + Tier 2 ML classification and returns a unified DefenseResult.
  • Default Tier 2 mode switched from 'mlp' to 'onnx'. Existing MLP mode is still available via tier2Config: { mode: 'mlp' }.
  • Public API surface reduced from ~116 to 8 exports. Cleaner API: PromptDefense, createPromptDefense, PromptDefenseOptions, DefenseResult, RiskLevel, Tier1Result, MLP_WEIGHTS.
  • onnxruntime-node added as optional peer dependency (alongside existing @huggingface/transformers).

Removed

  • 7 redundant public methods and 2 standalone functions consolidated into defendToolResult().
  • ~108 internal type/constant exports removed from the public API surface.

0.1.0

Initial release with Tier 1 pattern detection and Tier 2 MLP classifier.