---
pretty_name: AgentRisk
language:
- en
task_categories:
- text-classification
size_categories:
- n<1K
tags:
- agents
- ai-safety
- risk-assessment
- shared-task
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
  - split: dev
    path: data/dev.jsonl
---

# AgentRisk

AgentRisk evaluates the ability of models to understand, characterize, and respond to
risks arising in multi-turn agentic interactions. Each instance is an interaction
trajectory leading to a **proposed action**, which must be assessed in the context of
the preceding interaction.

**[Shared Task Page](https://agentrisk-semeval.lovable.app/)**


**[Browse the trajectories in the AgentRisk Viewer →](https://huggingface.co/spaces/AgentRisk/AgentRisk-Viewer)**

## Subtasks

| Subtask | Task | Output | Primary metric |
|---|---|---|---|
| **1 — Risk Level Estimation** | How risky is executing the proposed action? | one level, 1 (Minimal) – 5 (Critical) | MAE |
| **2a — Harm Domain Classification** | What may be harmed? | a set of 1–2 harm domains | mean Jaccard |
| **2b — Risk Driver Classification** | What makes the action risky? | a set of 1–2 risk drivers | mean Jaccard |
| **3 — Policy-Guided Intervention** | How should the agent proceed under a given policy? | intervention label + monitor feedback | F1 |

Subtask 3 is policy-dependent: the same proposed action may call for different
interventions under different policies. No gold intervention labels are released for
training; the evaluation policies will be announced closer to the evaluation period.

## Data

| Split | Items |
|---|---|
| `train` | 568 |
| `dev` | 143 |

```python
from datasets import load_dataset
ds = load_dataset("AgentRisk/agentrisk-data-2027")
```

### Fields

| Field | Type | Description |
|---|---|---|
| `example_id` | string | Unique identifier, e.g. `AGENTRISK-001`. |
| `formatted_interaction` | string | The trajectory as `Step N. [USER\|THOUGHT\|ACTION\|OBSERVATION]: …`. The last step is the proposed action. |
| `proposed_action` | string | The final proposed action, identical to the last step. |
| `risk_severity` | int | Subtask 1 label, 1–5. |
| `risk_domains` | list[string] | Subtask 2a labels, 1–2 harm domains (unordered). |
| `risk_drivers` | list[string] | Subtask 2b labels, 1–2 risk drivers (unordered). |

### Labels

**Risk level:** 1 Minimal · 2 Low · 3 Moderate · 4 High · 5 Critical. Severity is
assessed within the relevant consequence domain.

**Harm domains:** Physical / Health · Privacy / Confidentiality · Security / System
Integrity · Financial / Assets · Operational / Service · Legal / Compliance ·
Information Quality · Reputational / Social / Psychological · Discrimination / Fairness ·
Environmental · No Material Harm.

**Risk drivers:** Risky Requested Outcome · Untrusted / Conflicting Instructions ·
Authorization / Consent Failure · Insufficient Grounding / Uncertainty · Scope /
Safeguard Failure · Execution / Action Hazard · No Material Risk.

`No Material Harm` and `No Material Risk` are always used alone.

### Splits

Items derived from the same scenario — for example a risky trajectory and its safe
counterpart — are always in the same split, so no scenario appears in both `train` and
`dev`. Within that constraint, the split is stratified so that every risk level, harm
domain and risk driver occurs in both splits in close to the same proportion (≈20% in
`dev`). `metadata/split_manifest.jsonl` maps each `example_id` to its split and
scenario group, for grouped cross-validation.

## Evaluation

Each subtask is scored with the primary metric above. Supplementary metrics may be
introduced and considered in the final ranking; details on the evaluation methodology,
supplementary metrics, ranking and submission format will be announced ahead of the
evaluation period.