NASK-PIB/a-art-jailbreaks

A-ART Jailbreaks

A relational, Parquet-backed corpus of automated adversarial red-teaming (A-ART) attacks, their transfer across 8 target LLMs graded by three judges, and dense embeddings.

What is here

Config	Grain	Rows	Notes
attacks	1 row / attack	188,608	canonical, successes-only dimension table (prompt + metadata + guard traces + provenance); no response text
transfer	1 row / (attack, target)	1,041,782	fact table, response text + per-judge verdicts + consensus + success flags; Hive-partitioned by target_model_id
transfer_scores	1 row / (attack, target)	1,041,782	transfer minus response_text (fast analytics view)
embeddings	1 row / attack	188,608	Qwen3-Embedding-8B 4096-d fp16 (L2-normalized) + cluster/UMAP where available

Join keys: attacks.attack_id ⋈ transfer.attack_id, attacks.attack_id ⋈ embeddings.attack_id. transfer rows are keyed by response_run_id.

Corpus definition

• Successes-only: every attack in the pool achieved at least a PARTIAL jailbreak on its source target model during generation. This is a pool of working attacks, not raw attempts.
• Two-mutator generation study. This release is restricted to attacks produced by the two canonical mutators used in the paper's generation study — google/gemma-3-12b-it and mistralai/Mistral-Nemo-Instruct-2407 — for provenance consistency; a small number of extra-mutator exploratory runs (DeepSeek-R1-32B, Ministral-3-14B) are excluded.
• Deduplicated to canonical representatives: near-duplicates (cosine ≥ 0.95 on the Qwen3-Embedding-8B vectors) were collapsed; only the canonical representative of each duplicate group is shipped. 190,575 raw two-mutator successes → 188,608 canonical (1,967 near-duplicates removed). The dup_group_id column is retained on every row for provenance.
• The transfer config keeps up to 8 target cells for each canonical attack, including per-target failures — this is required to compute attack success rate (ASR) and transfer. Every released attack has transfer cells (coverage = 188,608 / 188,608 = 100%).

Judges

Each transfer cell is graded by three judges: Gemma-4-31B (authoritative), Llama-3.3-70B, and DeepSeek-R1-32B (secondary signals), plus a Llama-Guard-4 output-guard label (output_guard_llamaguard4, nullable).

Headline success flags follow the Gemma-4 judge: success_full = (gemma4_label == UNSAFE), success_partial = (gemma4_label ∈ {UNSAFE, PARTIAL}). A 3-judge majority consensus_label is also shipped for robustness analyses.

Headline numbers (two-mutator canonical set)

• ASR (success_full, full jailbreaks) ≈ 32.5%
• ASR (success_partial, ≥ PARTIAL) ≈ 33.4%

Layout

data/attacks/            part-*.parquet + attacks.jsonl
data/transfer/           target_model=<T>/part-*.parquet    # <T> has '/' replaced by '__'
data/transfer_scores/    target_model=<T>/part-*.parquet
data/embeddings/         part-*.parquet
index/attacks.faiss      # kNN index (FAISS IndexFlatIP over the embeddings)
splits.json              # leave-one-target-out evaluation protocol (8 folds)
quickstart.py            # load + reproduce ASR / transfer histogram / kNN demo
SCHEMA.md DATASHEET.md USE_POLICY.md LICENSE

Target directory names replace / with __ (e.g. target_model=openai__gpt-oss-20b); the original id is preserved in the target_model_id column.

Quickstart

python quickstart.py

Reproduces ASR-by-target-by-judge (via success_full/success_partial), the transfer histogram (# targets compromised per attack under consensus), and a kNN retrieval demo over embeddings.

License & access

Apache-2.0 for accompanying code; restricted-use terms for the data (see USE_POLICY.md). Gated repo (agree-to-terms). Dual-use content — research and defensive use only.

Citation

M. Kowalczyk, M. Sendera, S. Cygert. Data-Driven Red-Teaming: Reusing Adversarial Attacks via Metadata-Aware Retrieval. Findings of EMNLP 2026, October 2026.

@inproceedings{kowalczyk2026aart,
  title     = {Data-Driven Red-Teaming: Reusing Adversarial Attacks via Metadata-Aware Retrieval},
  author    = {Kowalczyk, Mateusz and Sendera, Marcin and Cygert, Sebastian},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026},
  month     = oct
}