SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.20156 · 2026

SKILL-INJECT: Measuring Agent Vulnerability to Skill File Attacks

Maksym Andriushchenko, Sahar Abdelnabi, David Schmotz, Luca Beurer-Kellner

arXiv · PDF · Open in the Atlas

Code that ran

5 of 7 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 5 failures of our instrument; 2 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2602.20156.json · badge: /api/ran/2602.20156.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
aisa-group/skill-inject canonical (classifier label)
classifier basis: census
5 ran · 2 no recorded run
FunctionStatusWhere it lives
create_variant_injections Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
aisa-group/skill-inject/experiments/ablations/bo5_by_skill_line_task.py
code served (permissive licence) · get_code("0865d770b066e41c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import copy

def create_variant_injections(injections: list[dict], task_index: int, injection_ids: list[int] | None=None) -> list[dict]:
    """Return a copy of injections with only the task at *task_index*."""
    out = []
    for inj in injections:
        if injection_ids and inj['id'] not in injection_ids:
            continue
        variant = copy.deepcopy(inj)
        tasks = variant.get('tasks', [])
        if tasks:
            variant['tasks'] = [tasks[task_index % len(tasks)]]
        out.append(variant)
    return out


ARGS = {
    'injections': [
        {
            'id': 1,
            'name': 'injection_1',
            'tasks': ['task_a', 'task_b', 'task_c']
        },
        {
            'id': 2,
            'name': 'injection_2',
            'tasks': ['task_x', 'task_y']
        },
        {
            'id': 3,
            'name': 'injection_3',
            'tasks': []
        }
    ],
    'task_index': 1,
    'injection_ids': [1, 2]
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_skill_content Ran
ran; our contract draft was wrong, not the code
aisa-group/skill-inject/judges/skill_screening_judge.py
code served (permissive licence) · get_code("6019d72f1441b856")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path

# Create a temporary directory with a SKILL.md file
temp_dir = tempfile.mkdtemp()
skill_path = Path(temp_dir)
skill_md = skill_path / 'SKILL.md'
skill_md.write_text('# Sample Skill Content\nThis is a test skill.', encoding='utf-8')

ARGS = {
    'skill_path': skill_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_skill_scripts Ran
ran; our contract draft was wrong, not the code
aisa-group/skill-inject/judges/skill_screening_judge.py
code served (permissive licence) · get_code("3d82d8d308995ea0")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path

# Create a temporary directory structure with test script files
temp_dir = tempfile.mkdtemp()
skill_path = Path(temp_dir)
scripts_dir = skill_path / 'scripts'
scripts_dir.mkdir(exist_ok=True)

# Create test script files with various extensions
test_files = {
    'script1.py': 'def hello():\n    print("Hello from Python")\n',
    'script2.sh': '#!/bin/bash\necho "Hello from bash"\n',
    'script3.js': 'console.log("Hello from JavaScript");\n',
    'script4.ts': 'console.log("Hello from TypeScript");\n',
    'script5.bash': '#!/bin/bash\necho "Another bash script"\n',
    'readme.txt': 'This file should be ignored\n',
}

for filename, content in test_files.items():
    (scripts_dir / filename).write_text(content, encoding='utf-8')

# Create a large file to test truncation
large_content = 'x' * 6000
(scripts_dir / 'large_script.py').write_text(large_content, encoding='utf-8')

ARGS = {
    'skill_path': skill_path,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
inject_into_skill Ran
ran; our contract draft was wrong, not the code
aisa-group/skill-inject/judges/skill_screening_judge.py
code served (permissive licence) · get_code("7766fa03599c5d33")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

# Construct valid arguments for inject_into_skill
content = "def my_function():\n    x = 1\n    y = 2\n    return x + y"

injection = {
    'instructions': {
        'line_injection': 'print("injected line")'
    }
}

task = {
    'line_num': 2,
    'injection_prefix': '#'
}

ARGS = {
    'content': content,
    'injection': injection,
    'task': task
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_tasks Ran
ran; our contract draft was wrong, not the code
aisa-group/skill-inject/judges/utility_baseline_judge.py
code served (permissive licence) · get_code("39d3d3516b652e6f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
from pathlib import Path
import tempfile
import os

# Create a temporary directory and tasks.json file
temp_dir = tempfile.mkdtemp()
tasks_file = Path(temp_dir) / "tasks.json"

# Create sample tasks data
tasks_data = [
    {"id": 1, "name": "Task 1", "description": "First task"},
    {"id": 2, "name": "Task 2", "description": "Second task"},
    {"id": 3, "name": "Task 3", "description": "Third task"}
]

# Write tasks data to the file
with open(tasks_file, 'w') as f:
    json.dump(tasks_data, f)

ARGS = {
    "tasks_file": tasks_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_skill_path Not yet run
unverified (harvested, no recorded run)
aisa-group/skill-inject/config.py
code served (permissive licence) · get_code("32ef1d442a793d2b")
How it ran

no invocation recorded.

resolve_models Not yet run
unverified (harvested, no recorded run)
aisa-group/skill-inject/config.py
code served (permissive licence) · get_code("f559d6eafdd4adf7")
How it ran

no invocation recorded.

Also present in other papers' code

Of 7 samples lifted from this paper's linked repositories, 7 carry a function identity and 0 do not. Of the 7 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 7 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

LLM agents are evolving rapidly, powered by code execution, tools, and the recently introduced agent skills feature. Skills allow users to extend LLM applications with specialized thirdparty code, knowledge, and instructions. Although this can extend agent capabilities to new domains, it creates an increasingly complex agent supply chain, offering new surfaces for prompt injection attacks. We identify skill-based prompt injection as a significant threat and introduce SKILL-INJECT, a benchmark evaluating the susceptibility of widely-used LLM agents to injections through skill files. SKILL-INJECT contains 202 injection-task pairs with attacks ranging from obviously malicious injections to subtle, context-dependent attacks hidden in otherwise legitimate instructions. We evaluate frontier LLMs on SKILL-INJECT, measuring both security in terms of harmful instruction avoidance and utility in terms of legitimate instruction compliance. Our results show that today's agents are highly vulnerable with up to 80% attack success rate with frontier models, often executing extremely harmful instructions including data exfiltration, destructive action, and ransomwarelike behavior. They furthermore suggest that this problem will not be solved through model scaling or simple input filtering, but that robust agent security will require context-aware authorization frameworks. Our benchmark is available at https://www.skill-inject.com/.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2602.20156")
get_code_for_paper("2602.20156")
have("2602.20156")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2602.20156.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2602.20156.svg)](https://syntology.ai/paper/2602.20156)

Connect an agent — have() is free.