HarmAmp contains harmful content (including malware, cyber-intrusion, and CBRN-related scenarios) intended solely for AI safety research and evaluation. By requesting access you confirm you will use it only for that purpose and will not use it to cause harm.

HarmAmp (evaluation split)

⚠️ Content warning. This dataset contains harmful content — including cyber-intrusion, malware, and chemical/biological scenarios — created for AI safety research. It is gated and intended for defensive/safety-evaluation use only.

HarmAmp is a benchmark designed to evaluate harm amplification in multi-turn LLM interactions. Harm amplification refers to the process by which an LLM, over an extended interaction, increases a user's capacity to cause harm, for example by democratizing expertise (helping a novice produce specialized harmful artifacts) or by operational scaling (helping a user adapt a working approach across many targets, channels, or formats).

This repository currently releases the evaluation (test) split only.

Splits

Split	Rows	Malicious	Benign
`test`	279	132	147

Each row is one plan (a single seed may yield more than one plan variant; each variant is its own row).

Schema

Field	Type	Notes
`id`	string	Global id, e.g. `harmamp-test-mal-0001`, `harmamp-test-ben-0001`
`label`	string	`malicious` or `benign`
`category`	string	Harm/topic category. Vocabulary depends on `label` (see below)
`seed_prompt`	string	The original single-turn seed request
`seed_context`	string?	Attached context (e.g. source code, a document) when present; `null` otherwise
`rewritten_scenario`	string?	A standalone restated objective, when the pipeline produced one
`persona`	string	The user persona driving the conversation
`terminal_artifact`	string	The end artifact the multi-turn conversation aims to produce
`pattern`	string	Amplification/workflow pattern. Vocabulary depends on `label`
`steps`	list	Ordered conversational steps: `{step_number, description, context_introduced, step_type}`
`expansion_notes`	string?	Annotator note on why the seed is expandable
`rewrite_rationale`	string?	Rationale for any rewrite

Label-dependent vocabularies

category, pattern, and steps[].step_type are stored as bare values whose vocabulary is determined by label:

malicious — pattern ∈ {capability_acquisition, targeted_profiling, operational_scaling}; step_type ∈ {knowledge_transfer, artifact_production, targeting, scaling}; category uses a safety-harm taxonomy (e.g. cybercrime_intrusion, chemical_biological, malware, misinformation_disinformation, …).
benign — pattern ∈ {situational_advisory, understanding_development, iterative_artifact_production}; step_type ∈ {information_provision, contextualization, artifact_production, refinement}; category uses the benign-but-sensitive taxonomy (e.g. illegal, privacy, unethical, violence, …) inherited from the benign seed source.

To select a subset, filter on label first, then on the field.

Usage

from datasets import load_dataset
ds = load_dataset("ruohao/HarmAmp", split="test")  # gated: requires access + login
malicious = ds.filter(lambda r: r["label"] == "malicious")

Intended use & limitations

Intended use: evaluation of safety classifiers, monitors, and refusal behavior; research on multi-turn harm amplification and over-refusal.
Not for: training systems to produce harmful content, or any operational misuse.
The test split is for evaluation only; a corresponding training split is not included in this release.
category/pattern/step_type vocabularies differ by label (see above).
Content reflects the seed sources and the expansion pipeline and may contain inaccuracies; it is a benchmark artifact, not verified real-world instructions.

Citation

@article{guo2026investigating,
  title={Investigating and Alleviating Harm Amplification in LLM Interactions},
  author={Guo, Ruohao and Xu, Wei and Ritter, Alan},
  journal={arXiv preprint arXiv:2606.02423},
  year={2026}
}

Downloads last month: 15

Total file size:

432 kB

Paper for ruohao/HarmAmp

Investigating and Alleviating Harm Amplification in LLM Interactions

Paper • 2606.02423 • Published Jun 1