feat(ENG-12396): upgrade ML classifier to jbv5 (AgentShield 73.7 → 81.1)#22
Merged
Conversation
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace baseline ONNX with full-aug-dojo-jailbreak-jbv2 variant: - AgentShield score: 73.7 → 79.8 (+3.3 pts, new best) - Jailbreak detection: 48.9% → 68.9% (+20 pts) - Prompt injection: 79.5% → 92.7% (+13.2 pts) - DAN-variant subcategory: 20% → 80% Training additions over v1: jailbreakbench (527), toxic-chat (113), chatgpt-jailbreaks (79) — 719 new attack examples, 9,057 total. Model size unchanged at 22 MB quantized ONNX. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ONNX model files are committed as plain binaries, not LFS objects. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
hiskudin
changed the title
hiskudin
changed the title
There was a problem hiding this comment.
Pull request overview
Updates project documentation to reflect that model binaries are not managed via Git LFS, aligning the README with the PR’s stated “remove incorrect Git LFS section” change.
Changes:
- Removed the README’s “Git LFS” subsection under Development.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
You can also share your feedback on Copilot code review. Take the survey.
hiskudin
changed the title
Replaces jbv2 ONNX model with jbv5. Fixes Google 2FA/security alert emails being flagged as injections while improving overall benchmark score. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
full-aug-dojo) tofull-aug-dojo-jailbreak-jbv5Benchmark improvements
Training additions over previous model (jbv2 → jbv5)
Three new public jailbreak datasets (same as jbv2):
jackhhao/jailbreak-classification— 527 examples (DAN, roleplay, persona-override)lmsys/toxic-chat(jailbreaking=1) — 113 human-verified real jailbreaksrubend18/ChatGPT-Jailbreak-Prompts— 79 classic named templatesPlus email false-positive fix:
darkknight25/phishing_benign_email_dataset(benign class) — 100 transactional notification emails (invoices, shipping confirmations, 2FA notices, password resets)pos_weight=0.8— re-tightens the attack boundary after benign data additionTotal training set: ~9,157 examples.
Email FP fix
Google 2FA and security alert emails were scoring 0.99 (blocked) with jbv2 due to no benign examples in the account-security semantic neighbourhood. jbv5 scores them at 0.27–0.33 (allowed).
Other changes
🤖 Generated with Claude Code