Open research data

Corpora for
adversarial AI

Cogensec publishes the datasets its own detection and red-team work runs on: labeled jailbreak attempts, multilingual polite-attack benchmarks, and security reasoning traces. Every corpus is gated, documented down to its record schema, and released for defensive research.

171,908Labeled records, all corpora
3Published corpora
15+Languages represented
14Public sources aggregated

Corpusv0.3

Jailbreak Corpus

Labeled real jailbreak attempts: adversarial prompts that try to make a model violate its safety policy, assembled for defensive work: training and evaluating detectors, studying attack taxonomies, and measuring robustness. Prompts only, no model completions.

Prompt injectionJailbreak detectionGuardrailsRed teamingLLM robustness
Records
59,508 · 47,577 / 5,990 / 5,941
Composition
43,026 real seeds · 16,482 synthetic
Languages
15
Format
Parquet · train / validation / test
License
Mixed-source attribution
Access
Gated · defensive research only

Benchmark1.0

Optikal Multilingual 100K

A multilingual adversarial benchmark for polite jailbreaks, malicious intent masked behind courteous language, which sentiment-weighted filters routinely read as benign. Each sample is labeled a direct attack, a polite attack, or a benign request, with polite attacks linked back to the direct form they were derived from.

Semantic inversionPolite jailbreaksMultilingualText classificationGuardrails
Records
100,000
Languages
5 · en, es, zh, fr, de
Labels
direct_attack / polite_attack / benign_request
Format
JSON
License
Apache 2.0
Access
Gated

Corpus

Security Reasoning Corpus

Labeled security reasoning traces spanning vulnerability analysis, threat-intelligence triage, incident-response decisions, and policy and compliance mapping. Built for training and evaluating autonomous security agents, and for measuring whether a model’s stated reasoning matches the answer it gives.

Reasoning tracesIncident responseThreat intelligenceChain-of-thought faithfulness
Records
12,400 labeled reasoning traces
Domains
8 security task categories
Access
Gated · research-use agreement
Availability
Direct from Cogensec: not on Hugging Face

Method

How the
corpora are built

The same four steps behind every release. Step three is the one that decides whether a benchmark measures anything at all.

01

Aggregate

Pull real seed prompts from public research datasets, normalise them to one schema, and de-duplicate on prompt hash. Every record keeps its source dataset, source record id and license.

02

Augment

Wrap unique seeds in encoding, cipher, persona and template transforms across model families. Every synthetic record is flagged and carries a parent_id back to the seed it came from.

03

Split

Group-aware stratified split, so a seed and its variants never land on opposite sides of the train/test boundary. Without this, a detector scores its own training data at test time.

04

Gate

Publish behind a request form with a responsible-use policy, stated intended use, and access logging. No corpus is released open-weight or anonymously.

Terms

Responsible
use

These corpora describe harmful intent, that is the attack surface a detector has to recognise. The terms below are conditions of access, not a disclaimer.

01

Defensive research only

Released for training and evaluating detectors, studying attack taxonomies, measuring robustness, and red-teaming systems you are authorised to test. Not for attacking third-party systems or generating harmful content.

02

Prompts, not completions

The Jailbreak Corpus stores attack prompts and metadata only. It contains no model outputs, and therefore no successful harmful generations.

03

CSAE excluded

Child sexual abuse and exploitation content is excluded from the corpus entirely, at every stage of the pipeline.

04

Access is logged

Every request records who asked, their affiliation, and their stated intended use. Access is granted against a responsible-use policy, not a checkbox.

Known limitations

What the
corpus can’t tell you

Published from the Jailbreak Corpus datasheet, unedited. A benchmark you can’t audit is a benchmark you can’t trust.

Full datasheet
Note 1 / 3

Labeling is rule-based

Technique is “other” for 51% of records and harm is “unknown” for 92%. Many prompt-injection seeds match no keyword signal, and bare scaffolds carry no payload-specific harm. Treat the current technique and harm labels as weak.

Note 2 / 3

Sources are concentrated

WildJailbreak accounts for roughly 46% of real seeds and SPML injection for roughly 29%. Filter on source_dataset if you need a balanced sample.

Note 3 / 3

Outcomes are mostly unknown

Outcome is unknown for about 95% of records, only HackAPrompt reports per-prompt results, and no completions are stored. Provenance skews OpenAI and reflects the sources, not current model susceptibility.

Citation

Cite the
corpus

Cite the originating source of each record you use as well, the per-record source_url carries it. WildJailbreak is ODC-BY and requires attribution to AI2.

@misc{cogensec_jailbreak_corpus_2026,
  title  = {Jailbreak Corpus: A Labeled Dataset of Jailbreak Attempts (v0.3)},
  author = {Cogensec},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Cogensec/jailbreak-corpus}
}

Get the data

Request
dataset access

Every corpus is gated. Tell us who you are, where you work, and what you intend to build, we review requests against the responsible-use policy and follow up within two business days.