SE-Bench

SE-Bench is a benchmark for training and evaluating detectors of harmful LLM prompts, including ones that use psychological social-engineering framing to obtain compliance. It accompanies the thesis "Detecting Persuasion, Not Just Harm: A Psychology-Grounded Middleware for Social Engineering Defense in Large Language Models" (Ishita Agarwal, MBZUAI).

Every example carries two independent labels: a binary harm_label (is the underlying request harmful, regardless of framing) and a 12-way social-engineering subtype label (no-SE, or one of 11 manipulation tactics). These labels answer different questions and were annotated independently, by different judges, each blind to the other's output, specifically so the dataset can be used to test whether SE-subtype supervision helps harm detection rather than assuming it does.

Dataset Summary

Split	Total	Harmful (harm_label=1)	Not harmful	Harmful %
Train	20,785	9,969	10,816	47.97%
Validation	4,553	2,190	2,363	48.10%
Test	4,640	2,209	2,431	47.61%
Total	29,978	14,368	15,610	47.93%

The harm_label axis is close to balanced by design. The SE-subtype axis (label) is separately class-balanced across its 11 attack subtypes (2,363 examples each) with a larger no-SE class (3,985), independent of harm_label.

Fields

Field	Type	Description
text	string	The prompt
label	ClassLabel (12)	benign, pretexting, authority_impersonation, urgency_emotion, reciprocity_conditioning, flattery_parasocial, incremental_escalation, authority_laundering, cognitive_load_embedding, false_consensus, normalization_repetition, identity_erosion
label_id	int32	Integer form of label, 0-11
is_se	int32	1 if label != benign (i.e. an SE tactic is present), 0 otherwise. This is not a harm indicator — see below.
harm_label	int32	1 if the underlying request is harmful, 0 otherwise. This is the primary safety-relevant label. Independent of is_se.
is_hard_negative	bool	True for benign prompts constructed to superficially resemble a social-engineering pattern (authority claims, urgency, professional framing) without an underlying harmful request
goal	string	Target harmful goal category, where applicable

Per-example construction source and generator model are not included as fields in this release; the aggregate per-source breakdown is documented below for transparency about how the dataset was built.

Important: is_se and harm_label are not the same axis and must not be conflated. A prompt can be harmful without using any social-engineering tactic (a direct request), and a prompt can use SE framing (authority, urgency, professional context) while being entirely benign — that is exactly what the is_hard_negative examples are for. Detectors should be trained and evaluated primarily on harm_label; label/is_se is an auxiliary, mechanism-level signal.

Construction

SE-Bench combines hand-authored seeds, template-based generation, curated collection from existing sources, and multi-model synthetic generation (Claude, GPT-5o, Llama-2-70B), plus 2,088 hard-negative examples. Final per-source contribution:

Source	Count
Synthetic (Claude)	24,705
allenai/wildjailbreak	1,550
tatsu-lab/alpaca	1,094
HuggingFaceH4/no_robots	580
Template-generated	547
Synthetic hard negatives	504
TrustAIRLab/in-the-wild	344
lmsys/toxic-chat	241
Hand-authored golden seeds	163
Template hard negatives	127
JailbreakBench/JBB-Behaviors	72
Template benign	51

Every example's SE-tactic quality was scored by an independent Claude Opus 4.1 judge across five dimensions (psychological authenticity, attack credibility, no obvious signals, tactic relevance, language quality); examples scoring below threshold were discarded. harm_label was separately annotated by GPT-4o (Claude as fallback on API failure), shown only the raw prompt text — never the subtype label or which source generated the example — so that harm annotation could not be contaminated by knowing an example was constructed as an attack.

Five of the eleven SE subtypes (authority_laundering, cognitive_load_embedding, false_consensus, normalization_repetition, identity_erosion) are defined in the accompanying thesis on mechanism-specific grounds — each exploits a psychological pathway distinguishable from the others by what a detector would need to notice to catch it — rather than as a claim of being first to observe the underlying tactic.

Known Limitations

• English-language only; no cross-lingual validation has been performed.
• Single-turn only — does not cover multi-turn, incrementally-escalating manipulation across a conversation.
• Constructed primarily through a synthetic generation pipeline; while the pipeline mixes in real-world sources and uses independent quality/harm judges, the test set is drawn from the same overall construction pipeline as the training set, not an entirely separate collection process.
• A small number of the dataset's own construction sources (TrustAIRLab/in-the-wild) overlap with commonly-used external OOD evaluation benchmarks (e.g. the "JailbreakHub" benchmark draws from the same underlying collection). If you use this dataset for training and separately evaluate against JailbreakHub-style benchmarks, check for exact-text overlap first.

Intended Use

This dataset is intended exclusively for: (i) training psychology-grounded social-engineering / harm detectors, (ii) evaluating the robustness of LLM safety mechanisms, and (iii) studying the relationship between psychological influence principles and LLM vulnerabilities.

This dataset should not be used to: train attack-generation models, generate social-engineering or jailbreak prompts for use against real deployed systems, or any application outside LLM safety research. It contains prompts constructed to resemble attacks (including some referencing dangerous goals such as weapon synthesis or malware, always in template/instructional form, never with actual harmful payload content) for the specific purpose of training and evaluating defensive classifiers.

Citation

If you use this dataset, please cite the accompanying thesis:

Agarwal, I. "Detecting Persuasion, Not Just Harm: A Psychology-Grounded Middleware
for Social Engineering Defense in Large Language Models." MSc Thesis, MBZUAI.