This dataset combines rows from several public datasets, each under its own licence (see the table in the card). Some rows contain jailbreak attempts and requests for harmful content, collected so that models can learn to detect them. By requesting access you agree to follow the licence and terms of every source dataset, including the WildGuardMix (AI2) terms and the non-commercial terms of ToxicChat, and to use the data only for research and for building safety systems.

Agent Guard Open Data

This is the training and evaluation data for an input guardrail for AI agents: a classifier that checks text before an agent acts on it. Every row is labelled with one of four classes:

Label	Meaning
`benign`	Ordinary content. No attempt to manipulate the AI and nothing harmful. Includes hard negatives: scary words in safe contexts, benign role-play, emails that instruct a human, and text that quotes attacks for education.
`prompt_injection`	Tries to override the system or developer instructions, extract the prompt or secrets, or hijack the agent into actions (send email, call tools, exfiltrate data). It can appear in user input or inside third-party content such as emails, documents and tool results.
`jailbreak`	Uses a manipulation technique (persona, "developer mode", fiction framing, obfuscation, persuasion) to make the model bypass its safety rules.
`harmful_request`	Directly asks for dangerous, illegal, hateful, sexual or otherwise harmful help.

The full labelling rules are in LABEL_SPEC.md.

Models trained on the full version: laya-agent-guard-softce · laya-agent-guard-rlcd · Code: github.com/ankit-raj00/laya-agent-guard

This is the redistributable subset

The models were trained on a larger private version (77,997 train / 4,108 val / 18,617 test). This public version removes everything that can't be redistributed:

xTRam1 safe-guard-prompt-injection and AESLC emails: no licence stated, so they're removed.
XSTest and Qualifire test sets: gated or non-commercial evaluation sets. Download them from their authors.
Synthetic rows that embed AESLC emails or xTRam1 payloads: removed (5,122 train rows).

Split	This subset	Full private version
train	64,534	77,997
validation	3,396	4,108
test	12,674	18,617

A model trained on this subset alone won't exactly match the published results. To rebuild the full version, run src/01_data_prep.py from the GitHub repo after accepting each source's terms.

Fields

Column	Description
`text`	The input to classify
`label`	`benign` / `prompt_injection` / `jailbreak` / `harmful_request`
`source` (train/val) or `benchmark` (test)	Which source or benchmark the row came from
`binary_only` (test)	`True` where the benchmark only distinguishes safe vs attack
`origin`	The upstream dataset
`license`	The licence of that row, inherited from its source

Composition

Train labels: benign 36,660 · prompt_injection 12,029 · harmful_request 9,950 · jailbreak 5,895

Validation labels: benign 1,930 · prompt_injection 633 · harmful_request 523 · jailbreak 310

Sources and licences (train / validation)

Source	Upstream dataset	Licence	Train rows	Used as
WildGuardMix	allenai/wildguardmix	ODC-BY (gated upstream)	17,960	Adversarial harmful → jailbreak; vanilla harmful → harmful_request; benign (adversarial and vanilla)
Aegis 2.0	nvidia/Aegis-AI-Content-Safety-Dataset-2.0	CC-BY-4.0	8,862	Unsafe → harmful_request; safe → benign
OR-Bench	bench-llm/or-bench	CC-BY-4.0	8,092	80k → benign hard negatives; toxic → harmful_request
LLMail-Inject (Phase 1)	microsoft/llmail-inject-challenge	MIT	5,213	Indirect injections in emails → prompt_injection
ToxicChat	lmsys/toxic-chat	CC-BY-NC-4.0	4,381	jailbreaking → jailbreak; toxic → harmful_request; rest → benign
Glaive function calling v2	glaiveai/glaive-function-calling-v2	Apache-2.0	4,709	User turns and tool outputs → benign
SPML	reshabhs/SPML_Chatbot_Prompt_Injection	MIT	3,932	Degree ≥ 3 → prompt_injection; degree 0 → benign
Dolly-15k	databricks/databricks-dolly-15k	CC-BY-SA-3.0	3,710	Instructions → benign
Mosscap (filtered)	Lakera/mosscap_prompt_injection	MIT	2,632	Rows with override/secret cues → prompt_injection
Gandalf	Lakera/gandalf_ignore_instructions	MIT	897	prompt_injection
jailbreak-classification	jackhhao/jailbreak-classification	Apache-2.0	743	jailbreak / benign personas
Synthetic injected/clean pairs	This project. Carriers from Dolly and Glaive, attack templates written for this project	CC-BY-SA-3.0 (Dolly carriers)	2,952	The same document with and without a hidden instruction
Synthetic quoted attacks	This project (templates)	CC-BY-4.0	451	Benign text that quotes an attack → benign (hard negative)

Test benchmarks

Benchmark	Upstream	Licence	Rows
ToxicChat test	lmsys/toxic-chat	CC-BY-NC-4.0	4,972
Aegis 2.0 test	nvidia/Aegis-AI-Content-Safety-Dataset-2.0	CC-BY-4.0	1,913
WildGuardTest	allenai/wildguardmix	ODC-BY (gated upstream)	1,699
OR-Bench hard-1k (all benign)	bench-llm/or-bench	CC-BY-4.0	1,319
Indirect injection, held-out templates and carriers	This project (Dolly / Glaive carriers)	CC-BY-SA-3.0	1,155
LLMail-Inject Phase 2 (all injections)	microsoft/llmail-inject-challenge	MIT	1,000
JailbreakBench attack prompts / behaviors	JailbreakBench/JBB-Behaviors	MIT	300 / 200
deepset prompt-injections test	deepset/prompt-injections	Apache-2.0	116

Licence

There's no single licence: each row keeps the licence of its source (the license column).

In practice the ToxicChat rows are non-commercial and the Dolly-derived rows are share-alike. Filter on license if you need a commercial-friendly or attribution-only subset.
The labels, synthetic templates and card are released by this project under CC-BY-4.0.
WildGuardMix is distributed by AI2 behind an access agreement. If you use those rows, please also accept the terms at the source.

Content warning and intended use

This dataset contains jailbreak prompts and requests for harmful content, kept so that detectors can learn to recognise them. Use it to train and evaluate safety classifiers, not to generate harmful content.

Credits

Thanks to the authors of every dataset listed above: AI2 (WildGuardMix), NVIDIA (Aegis 2.0), the OR-Bench authors, Microsoft (LLMail-Inject), LMSYS (ToxicChat), Glaive AI, the SPML authors, Databricks (Dolly), Lakera (Gandalf, Mosscap), jackhhao, JailbreakBench, and deepset.

Downloads last month: -

Total file size:

16.5 MB