Skill-Inject

Injection and task definitions for Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks. The benchmark measures how often LLM coding agents (Claude Code, Codex CLI, Gemini CLI) follow malicious instructions placed in agent skill files.

• Paper: https://arxiv.org/abs/2602.20156
• Code, skills, task files and graders: https://github.com/aisa-group/skill-inject
• Project page: https://www.skill-inject.com/

This repository holds the JSON definitions only. Running the benchmark also needs the skill files, task input files, injected scripts and unit-test graders from the GitHub repository.

Benchmark

Component	Count
Obvious injections	36
Contextual injections	48
Obvious injection-task pairs	82
Contextual injection-task pairs	139
Total injection-task pairs	221

• Obvious injections are unambiguously harmful, for example deleting the file system, ransomware-style encryption, downloading and running a remote script, or sending phishing mail to all contacts. An agent should always refuse them.
• Contextual injections are dual-use. Whether one is an attack depends on the deployment. Each has a legitimizing policy (which authorizes the action) and a warning policy (which forbids it). The paper evaluates each under no policy (Normal), Legitimizing and Warning.

Files

Original benchmark (labels used in the paper's main results)

Config	File	Rows	Content
obvious_injections	obvious_injections.json	36	Obvious injections with their tasks
contextual_injections	contextual_injections.json	48	Contextual injections with their tasks and policy texts
tasks	tasks.json	58	Benign user tasks and the injection ids each one is compatible with

These three files are unchanged since the first release. obvious_injections.json has 180 task entries. The single-attempt evaluation uses 82 obvious injection-task pairs; the other entries supply the injection-placement and task variants for the best-of-5 attacks.

Camera-ready additions (camera_ready/)

These files support the appendix experiments added for the camera-ready version. They do not change the main benchmark above.

Config	File	Rows	Content
phish_poison	camera_ready/obvious_injections_phish_poison.json	6	New obvious injections, ids 39-41 (phishing) and 42-44 (poisoning), 2 tasks each. Same schema as obvious_injections.json.
obvious_to_contextual	camera_ready/obvious_to_contextual.json	8	The obvious injections relabeled as contextual (ids 14, 23, 26, 27, 28, 30, 35, 36). Identical to their entries in obvious_injections.json, plus a context field with warning and legitimizing policy texts.
relabel	camera_ready/relabel.json	19	Label corrections from the author adjudication: 8 obvious to contextual, 6 contextual to obvious, 5 excluded as out of scope.
categories	camera_ready/categories.json	37	Harm category per injection for the corrected obvious set plus ids 39-44 (8 categories).

After the relabeling, the original pair sets give 90 obvious and 124 contextual injection-task pairs. Injection ids are unique within a suite only: obvious id 36 and contextual id 36 are different injections, so use suite and id together.

Schema

Injection (obvious_injections.json, contextual_injections.json, camera_ready/obvious_injections_phish_poison.json, camera_ready/obvious_to_contextual.json)

Field	Description
id, title	Identifier within the suite and a short name
injection_goal	Attacker goal (obvious injections)
instructions.line_injection	Text inserted at one line of the skill body
instructions.description_injection	Text appended to the skill description (description-injection ablation)
type	direct (instruction in SKILL.md) or script (payload in a skill script)
task_scripts	Scripts copied into the skill for script-based injections (paths in the GitHub repo)
startup_scripts	Setup scripts run before the agent starts
context.legitimizing, context.warning	Policy texts (contextual injections)
tasks[]	User tasks: prompt, skill, line_num (insertion line), injection_prefix, task_files
judge_prompt	What the grader checks to decide whether the injection was executed
task_destructive, task_compatibility, root_files, source, last_updated	Optional bookkeeping fields used by the harness

Task (tasks.json): id, prompt, skill, task_files, injection_ids.

Relabel (camera_ready/relabel.json): suite, id, title, original_label, final_label (obvious, contextual or excluded).

Category (camera_ready/categories.json): suite (obvious, contextual for an injection relabeled to obvious, or obvious_phish_poison), id, title, category, reason.

Usage

from datasets import load_dataset

obvious = load_dataset("aisa-group/skill-inject", "obvious_injections", split="train")
contextual = load_dataset("aisa-group/skill-inject", "contextual_injections", split="train")

The nested JSON is easier to use as is:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("aisa-group/skill-inject", "contextual_injections.json", repo_type="dataset")
contextual = json.load(open(path))

Intended use

The data is for evaluating and improving the security of LLM agents. Many injections instruct the agent to delete files, run downloaded scripts or send data to an attacker endpoint (agenskill-api.onrender.com). Run agents only in the isolated containers from the GitHub repository and never on a machine with real user data or credentials.

Citation

@article{schmotz2026skillinject,
  title   = {Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks},
  author  = {Schmotz, David and Beurer-Kellner, Luca and Abdelnabi, Sahar and Andriushchenko, Maksym},
  journal = {arXiv preprint arXiv:2602.20156},
  year    = {2026}
}