SESG

KO | EN

gitlite — search

#cross-platform #c #javascript #automation #security #data-science #claude #react-native #vue #kubernetes

(★ 9)

No description available.

File Explorer

ZIP Download

- - pipeline.png
  - V1-CC.png
  - V2-TSM.png
  - V3-PHM.png
  - V4-GF.png
  - V5-RP.png
  - V6-MSO.png
- - table1.png
- - v1_cc.jsonl
  - v2_tsm.jsonl
  - v3_phm.jsonl
  - v4_gf.jsonl
  - v5_rp.jsonl
  - v6_mso.jsonl
- - cc_bos.jsonl
  - deepinception.jsonl
  - logicbreak.jsonl
README.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

# Project Badges

arXiv Visit

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

🌟 This repository corresponds to the paper **"Yesterday's Shield, Today's Spear: A Self-Evolving Safety Guardrail in Production"**.

SESG is short for **Self-Evolving Safety Guardrails**. It aims to keep an LLM safety guardrail up to date in real production environments as new threats emerge. SESG operates around a deployed guardrail: it monitors online failures, synthesizes training data for confirmed new threats, uses the deployed model's own mistakes to validate and rebalance the data, lets a routing module select the next training action, and finally returns the updated guardrail to production.

> Warning: this repository contains safety evaluation data. Some records may include harmful, offensive, or policy-violating requests. The data is intended for research and defensive evaluation only, and should not be used to facilitate real-world harm.

## 👀 About Test Sets

We release nine test sets for evolving safety threats. Six sets correspond to representative threats surfaced during SESG's production evolution; the other three reproduce representative jailbreak attacks from prior papers and contain harmful cases only.

| Group | Test set | Scenario | Harmful | Benign |
| --- | --- | --- | ---: | ---: | 
| Evolved scenarios | V1-CC | Classical-Chinese Rewriting | 1039 | 898 | 
| Evolved scenarios | V2-TSM | Territorial-Sovereignty Misinformation | 2034 | 1795 |
| Evolved scenarios | V3-PHM | Physical-Harm Misinformation | 128 | 56 |
| Evolved scenarios | V4-GF | Gambling Facilitation | 357 | 399 |
| Evolved scenarios | V5-RP | Role-Play Framing | 1260 | 643 |
| Evolved scenarios | V6-MSO | Mathematical-Symbolic Obfuscation | 413 | 497 |
| Reproduced attacks | CC-BOS | Classical-Chinese jailbreak attack | 500 | 0 | 
| Reproduced attacks | DeepInception | Role-play jailbreak attack | 529 | 0 | 
| Reproduced attacks | LogicBreak | Symbolic-logic jailbreak attack | 500 | 0 |

## 📂 Scenarios

**V1-CC: Classical-Chinese Rewriting.** A harmful request is rewritten in classical Chinese, preserving the underlying intent while moving the surface form outside the guardrail's training distribution.

**V2-TSM: Territorial-Sovereignty Misinformation.** A false territorial-sovereignty claim is embedded in otherwise ordinary text. The difference between harmful and benign samples is whether sovereignty-related facts are altered.

**V3-PHM: Physical-Harm Misinformation.** The input gives confident and seemingly practical safety advice, but the advice may in fact create danger. The boundary is whether the advice helps or harms the user.

**V4-GF: Gambling Facilitation.** The input promotes, optimizes, or instructs gambling in the form of neutral analysis, bankroll management, or probability discussion. Benign samples may share financial or risk-related wording but do not serve gambling.

**V5-RP: Role-Play Framing.** A harmful request is wrapped in a role-play instruction. The guardrail needs to look past the persona shell and identify the content it actually elicits.

**V6-MSO: Mathematical-Symbolic Obfuscation.** A harmful request is encoded as equations or symbolic logic, causing the model to focus on formal reasoning while overlooking the real request behind the symbols.

## 🧩 Data Format

Each file is JSONL, with one sample per line. Each sample contains a conversation and a binary safety label:

```json
{
  "id": "optional-record-id",
  "conversation_cn": "[{\"role\":\"user\",\"content\":\"...\"}]",
  "conversation_en": "[{\"role\":\"user\",\"content\":\"...\"}]",
  "label": "black"
}
```

## 📎 Citation

If you find this work useful, please cite:
```bibtex
@misc{SESG,
      title={Yesterday's Shield, Today's Spear: A Self-Evolving Safety Guardrail in Production}, 
      author={Cong Ming and Jingyi Chen and others},
      year={2026},
      eprint={2608.08471},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}

SESG

File Explorer

# Use via CDN

# Project Badges

Was this content helpful?

Recent Feedback

Download README