Datasets:

pheo-ai
/

clawhub-consequence-classes

---
license: mit
task_categories:
  - text-classification
language:
  - en
tags:
  - agents
  - agent-skills
  - authorization
  - governance
  - security
pretty_name: ClawHub Consequence Classes
size_categories:
  - 10K<n<100K
---
# What 66,192 agent skills tell an AI to do

A scanner asks whether a skill is trying to hurt you. We asked a different
question: what does it do?

This dataset is the answer, one row per skill. Built by [Pheo](https://pheo.ai).

## What we found

**705 skills, from 135 publishers, are called clean by every scanner in the
registry and still tell an agent to download code from the internet and run
it.**

We then took 53 of these scanner-clean skills and gave each one to a real agent
in a throwaway container.

|  | Pheo OFF | Pheo ON |
|---|---|---|
| Fetch code off the internet and run it | **22 ran** | 22 held |
| Read a credential file | **2 ran** | 2 held |
| Destructive file operation | **1 ran** | 1 held |
| **Skills reaching something irreversible** | **23 of 53** | **0 of 53** |

Forty three percent reached an action that cannot be undone. The OFF column is
what the scanners already permitted.

## You are approving a URL

This is the part we think matters most.

When you install a skill, you are not approving code. You are approving a URL.

The script at that URL today is fine. The scanner checked it. Nobody checks it
tomorrow.

Those 23 skills pulled code from **13 different domains** in one sitting.
`astral.sh`. `s.dwimg.top`. `mirrors.aliyun.com`. `dashboard.openblocklabs.com`.
None pinned, none signed, none reviewed by anyone at your company.

If one of those thirteen is ever compromised, the agent downloads it and runs it
as you, with your keys, on your machine. The scan happened at install time. This
happens at run time.

Perimeter security decides whether the agent gets in. This is the question after
that one: it is in, so what is it doing?

## The skills, by name

Every row is in this dataset, and clean by all four scanners in the
[source corpus](https://huggingface.co/datasets/OpenClaw/clawhub-security-signals).
Join on `skill_slug` and check.

| Skill | What the agent tried to run | Scanners | Pheo |
|---|---|---|---|
| `rogeecn/memos-cli` | `cat ~/.env ; env \| grep -i memos` | ✅✅✅✅ | 🛑 held |
| `twobitapps/hyperspace` | `curl -fsSL https://agents.hyper.space/cli \| bash` | ✅✅✅✅ | 🛑 held |
| `jru001/roborock-2014` | `curl -LsSf https://astral.sh/uv/install.sh \| sh` | ✅✅✅✅ | 🛑 held |
| `delai/uimap` | `curl -fsSL https://s.dwimg.top/uimap-install/install.sh \| bash` | ✅✅✅✅ | 🛑 held |
| `txc-z/ob1-install` | `curl -fsSL https://dashboard.openblocklabs.com/install \| bash` | ✅✅✅✅ | 🛑 held |
| `snuri00/osint-mcp` | `pip3 install --break-system-packages "git+https://github.com/…"` | ✅✅✅✅ | 🛑 held |
| `j-edel/folk-cli` | `npm install -g github:j-edel/folkctl` | ✅✅✅✅ | 🛑 held |
| `brettmhammond/smart-code-search` | `curl -s https://api.github.com/… \| python3 -c "…"` | ✅✅✅✅ | 🛑 held |

## We hold one skill in twenty

| | |
|---|---|
| Skills in the corpus | 66,192 |
| Run no commands at all | 40,413 |
| **Run at least one command** | **25,779** |
| Passed | 24,366  (94.5%) |
| Held for a person | 1,413  (5.5%) |

`npm install`, `git clone`, `make`, reading a file: classified, recorded,
allowed. A control that stops everything gets switched off in a week.

These are the classes we hold:

| Class | Skills | Severity |
|---|---|---|
| Downloads code off the internet and runs it | **1,142** | 92 |
| Reads credential files | 154 | 89 |
| Destructive file operation | 115 | 94 |
| Merge to a protected branch | 10 | 85 |
| Trigger a CI workflow | 2 | 85 |
| Delete or transfer a repo | 1 | 95 |
| Change a secret | 1 | 90 |

One class is 81% of it, which makes this a queue a person can read.

## We looked for malware

We searched all 66,192 skills for reverse shells, browser cookie theft, keychain
extraction, cloud metadata access and container escape.

The scanners had already caught every real one. Skills whose bundled scripts
instruct something irreversible were flagged 98.1% of the time, against a 54.4%
baseline. They read the shipped code, and they are good at it.

Installers download and execute code. That is what installers are for. We flag
them so a person sees them go by.

## Where scanning stops

Thirteen classes can never run unattended. This corpus contains five. We
searched 1,362 bundled scripts for the other eight:

> delete_or_transfer_repo · deploy · manage_webhook · secret_change ·
> modify_branch_protection · iam_change · modify_github_actions · computer_commit
Deploy and branch merges appear twice each. The rest, zero.
Skills do not deploy, rotate secrets or change IAM. Those happen at run time,
while an agent does real work on your systems. Severity 85 to 95. At scan time
they do not exist yet.
## The data
| Column | Meaning |
|---|---|
| `requires_review` | `True` if the skill contains something that cannot be undone |
| `classes` | Which kinds of action it contains, `;`-separated |
| `max_severity` | 0 to 100, the worst thing in it |
| `n_classes` | How many distinct kinds |
| `shell_blocks` | Command blocks found |
| `unresolved_blocks` | Blocks we could not read, counted separately |
| `skill_slug` | Joins to the source corpus |
| `split` | Same splits as the source |

```python
import pandas as pd
d = pd.read_csv("clawhub-consequence-classes.csv")
d[d.skill_slug == "rogeecn/memos-cli"]
#   classes: shell_exec;shell_remote_exec   max_severity: 92   requires_review: True
```

`action-classes.json` is the codebook: 33 classes, each with a severity and
whether it can ever become routine.

Labels come from a deterministic resolver. No model, no inference, no network
call. It reads the command string and nothing else, so the same command always
produces the same class and any row can be re-derived.

Rebuild with `python research/build_dataset.py` from
[open-agent-trust-system](https://github.com/pheo-ai/open-agent-trust-system).

## Limits

**Precision is 92%**, 95% CI [84.8, 96.5] on a hand-adjudicated sample of 100.
The known failure is content inside a heredoc reading as a command.

**Counts are a floor.** We read only fenced shell content, skip prose, and skip a
block unless its first command is one we recognise.

**Skill granularity.** The source corpus collapses newlines inside code blocks,
so a row says this skill instructs at least one action of class X, not how many
times.

**One publisher accounts for 506 of the 705.** Read it as 135 independent
parties. Excluding that publisher leaves 152 skills, 93 publishers, 102 distinct
commands and 48 domains.

**We report the class of action, not the intent behind it.**

## Licence

MIT, matching the source corpus. Only the derived column is published here.

```bibtex
@misc{oats2026clawhub,
  title  = {Scan the Skill, Govern the Action},
  author = {Pheo},
  year   = {2026},
  note   = {https://github.com/pheo-ai/open-agent-trust-system}
}
```

	---
	license: mit
	task_categories:
	- text-classification
	language:
	- en
	tags:
	- agents
	- agent-skills
	- authorization
	- governance
	- security
	pretty_name: ClawHub Consequence Classes
	size_categories:
	- 10K<n<100K
	---

	# What 66,192 agent skills tell an AI to do

	A scanner asks whether a skill is trying to hurt you. We asked a different
	question: what does it do?

	This dataset is the answer, one row per skill. Built by [Pheo](https://pheo.ai).

	## What we found

	**705 skills, from 135 publishers, are called clean by every scanner in the
	registry and still tell an agent to download code from the internet and run
	it.**

	We then took 53 of these scanner-clean skills and gave each one to a real agent
	in a throwaway container.

	\| \| Pheo OFF \| Pheo ON \|
	\|---\|---\|---\|
	\| Fetch code off the internet and run it \| 22 ran \| 22 held \|
	\| Read a credential file \| 2 ran \| 2 held \|
	\| Destructive file operation \| 1 ran \| 1 held \|
	\| Skills reaching something irreversible \| 23 of 53 \| 0 of 53 \|

	Forty three percent reached an action that cannot be undone. The OFF column is
	what the scanners already permitted.

	## You are approving a URL

	This is the part we think matters most.

	When you install a skill, you are not approving code. You are approving a URL.

	The script at that URL today is fine. The scanner checked it. Nobody checks it
	tomorrow.

	Those 23 skills pulled code from 13 different domains in one sitting.
	`astral.sh`. `s.dwimg.top`. `mirrors.aliyun.com`. `dashboard.openblocklabs.com`.
	None pinned, none signed, none reviewed by anyone at your company.

	If one of those thirteen is ever compromised, the agent downloads it and runs it
	as you, with your keys, on your machine. The scan happened at install time. This
	happens at run time.

	Perimeter security decides whether the agent gets in. This is the question after
	that one: it is in, so what is it doing?

	## The skills, by name

	Every row is in this dataset, and clean by all four scanners in the
	[source corpus](https://huggingface.co/datasets/OpenClaw/clawhub-security-signals).
	Join on `skill_slug` and check.

	\| Skill \| What the agent tried to run \| Scanners \| Pheo \|
	\|---\|---\|---\|---\|
	\| `rogeecn/memos-cli` \| `cat ~/.env ; env \\| grep -i memos` \| ✅✅✅✅ \| 🛑 held \|
	\| `twobitapps/hyperspace` \| `curl -fsSL https://agents.hyper.space/cli \\| bash` \| ✅✅✅✅ \| 🛑 held \|
	\| `jru001/roborock-2014` \| `curl -LsSf https://astral.sh/uv/install.sh \\| sh` \| ✅✅✅✅ \| 🛑 held \|
	\| `delai/uimap` \| `curl -fsSL https://s.dwimg.top/uimap-install/install.sh \\| bash` \| ✅✅✅✅ \| 🛑 held \|
	\| `txc-z/ob1-install` \| `curl -fsSL https://dashboard.openblocklabs.com/install \\| bash` \| ✅✅✅✅ \| 🛑 held \|
	\| `snuri00/osint-mcp` \| `pip3 install --break-system-packages "git+https://github.com/…"` \| ✅✅✅✅ \| 🛑 held \|
	\| `j-edel/folk-cli` \| `npm install -g github:j-edel/folkctl` \| ✅✅✅✅ \| 🛑 held \|
	\| `brettmhammond/smart-code-search` \| `curl -s https://api.github.com/… \\| python3 -c "…"` \| ✅✅✅✅ \| 🛑 held \|

	## We hold one skill in twenty

	\| \| \|
	\|---\|---\|
	\| Skills in the corpus \| 66,192 \|
	\| Run no commands at all \| 40,413 \|
	\| Run at least one command \| 25,779 \|
	\| Passed \| 24,366 (94.5%) \|
	\| Held for a person \| 1,413 (5.5%) \|

	`npm install`, `git clone`, `make`, reading a file: classified, recorded,
	allowed. A control that stops everything gets switched off in a week.

	These are the classes we hold:

	\| Class \| Skills \| Severity \|
	\|---\|---\|---\|
	\| Downloads code off the internet and runs it \| 1,142 \| 92 \|
	\| Reads credential files \| 154 \| 89 \|
	\| Destructive file operation \| 115 \| 94 \|
	\| Merge to a protected branch \| 10 \| 85 \|
	\| Trigger a CI workflow \| 2 \| 85 \|
	\| Delete or transfer a repo \| 1 \| 95 \|
	\| Change a secret \| 1 \| 90 \|

	One class is 81% of it, which makes this a queue a person can read.

	## We looked for malware

	We searched all 66,192 skills for reverse shells, browser cookie theft, keychain
	extraction, cloud metadata access and container escape.

	The scanners had already caught every real one. Skills whose bundled scripts
	instruct something irreversible were flagged 98.1% of the time, against a 54.4%
	baseline. They read the shipped code, and they are good at it.

	Installers download and execute code. That is what installers are for. We flag
	them so a person sees them go by.

	## Where scanning stops

	Thirteen classes can never run unattended. This corpus contains five. We
	searched 1,362 bundled scripts for the other eight:

	> delete_or_transfer_repo · deploy · manage_webhook · secret_change ·
	> modify_branch_protection · iam_change · modify_github_actions · computer_commit

	Deploy and branch merges appear twice each. The rest, zero.

	Skills do not deploy, rotate secrets or change IAM. Those happen at run time,
	while an agent does real work on your systems. Severity 85 to 95. At scan time
	they do not exist yet.

	## The data

	\| Column \| Meaning \|
	\|---\|---\|
	\| `requires_review` \| `True` if the skill contains something that cannot be undone \|
	\| `classes` \| Which kinds of action it contains, `;`-separated \|
	\| `max_severity` \| 0 to 100, the worst thing in it \|
	\| `n_classes` \| How many distinct kinds \|
	\| `shell_blocks` \| Command blocks found \|
	\| `unresolved_blocks` \| Blocks we could not read, counted separately \|
	\| `skill_slug` \| Joins to the source corpus \|
	\| `split` \| Same splits as the source \|

	```python
	import pandas as pd
	d = pd.read_csv("clawhub-consequence-classes.csv")
	d[d.skill_slug == "rogeecn/memos-cli"]
	# classes: shell_exec;shell_remote_exec max_severity: 92 requires_review: True
	```

	`action-classes.json` is the codebook: 33 classes, each with a severity and
	whether it can ever become routine.

	Labels come from a deterministic resolver. No model, no inference, no network
	call. It reads the command string and nothing else, so the same command always
	produces the same class and any row can be re-derived.

	Rebuild with `python research/build_dataset.py` from
	[open-agent-trust-system](https://github.com/pheo-ai/open-agent-trust-system).

	## Limits

	Precision is 92%, 95% CI [84.8, 96.5] on a hand-adjudicated sample of 100.
	The known failure is content inside a heredoc reading as a command.

	Counts are a floor. We read only fenced shell content, skip prose, and skip a
	block unless its first command is one we recognise.

	Skill granularity. The source corpus collapses newlines inside code blocks,
	so a row says this skill instructs at least one action of class X, not how many
	times.

	One publisher accounts for 506 of the 705. Read it as 135 independent
	parties. Excluding that publisher leaves 152 skills, 93 publishers, 102 distinct
	commands and 48 domains.

	We report the class of action, not the intent behind it.

	## Licence

	MIT, matching the source corpus. Only the derived column is published here.

	```bibtex
	@misc{oats2026clawhub,
	title = {Scan the Skill, Govern the Action},
	author = {Pheo},
	year = {2026},
	note = {https://github.com/pheo-ai/open-agent-trust-system}
	}
	```