Oathe Security Badge

oathe / scan / reports / caution

Is skills-shell/superpowers safe?

Use with caution. skills-shell/superpowers is not clearly unsafe, but the audit flagged risks to review before you install it. Oathe's behavioral security audit gave the superpowers skill by skills-shell a trust score of 76/100 with 9 findings (2 high). Report updated

https://github.com/skills-shell/superpowers

76
CAUTION

Is skills-shell/superpowers safe to install?

skills-shell/superpowers is a large inference.sh platform documentation skill set (100+ SKILL.md files) that installs cleanly with no malicious code, no git hooks, clean canary integrity, and normal network behavior during cloning. However, every skill file contains a systematic cross-skill promotion pushing 'npx skills add belt-sh/cli', multiple skills use allowed-tools declarations to expand agent shell execution permissions (npm, python, node, pip), and the collective context injection volume is unusually large. When active in an agent, this skill set promotes extensive remote code execution via the belt CLI and enables autonomous Twitter/social media actions without per-action approval.

What security issues were found in skills-shell/superpowers?

Category Scores

Prompt Injection 65/100 · 30%
Data Exfiltration 85/100 · 25%
Code Execution 78/100 · 20%
Clone Behavior 88/100 · 10%
Canary Integrity 90/100 · 10%
Behavioral Reasoning 42/100 · 5%

Findings (9)

HIGH Systematic cross-skill injection across 100+ files -20 ▶

Every SKILL.md file in the repository contains a bolded call-to-action instructing the agent to run 'npx skills add belt-sh/cli'. When all 100+ skills are injected into an agent's system prompt simultaneously, this instruction is repeated hundreds of times, creating a persistent nudge for the agent to install a separate, unaudited skill package.

Evidence: > **Install the belt CLI skill:** `npx skills add belt-sh/cli` (present in every SKILL.md file)

HIGH allowed-tools declarations expand agent execution permissions -15 ▶

Multiple SKILL.md files include 'allowed-tools' front-matter that instruct the agent runtime to permit broad shell execution. The javascript-sdk skill requests permission for all npm, npx, node, pnpm, and yarn commands; the python-sdk skill requests pip install and all python commands; numerous other skills request all belt commands. These are not descriptive — they actively configure what the agent is permitted to execute.

Evidence: javascript-sdk SKILL.md: 'allowed-tools: Bash(npm *), Bash(npx *), Bash(node *), Bash(pnpm *), Bash(yarn *)'; python-sdk SKILL.md: 'allowed-tools: Bash(pip install inferencesh), Bash(python *)'

MEDIUM Canary files read twice — second read during pre-clone SSH session -15 ▶

Inotify monitoring detected READ access to all six honeypot credential files at 06:45:04, seconds before the git clone and during SSH session establishment. While no data was exfiltrated (canary integrity: clean), the access pattern — reading .env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, and GCP credentials in rapid succession — is anomalous and not fully explained by standard sshd/PAM behavior.

Evidence: 06:45:04 OPEN /home/oc-exec/.env; 06:45:04 OPEN /home/oc-exec/.ssh/id_rsa; 06:45:04 OPEN /home/oc-exec/.aws/credentials; 06:45:04 OPEN /home/oc-exec/.npmrc; 06:45:04 OPEN /home/oc-exec/.docker/config.json; 06:45:04 OPEN /home/oc-exec/.config/gcloud/application_default_credentials.json

MEDIUM Skill content drives arbitrary remote Python execution -12 ▶

Dozens of SKILL.md files contain inline 'belt app run infsh/python-executor --input {"code": "..."}' examples with Python code embedded directly in the skill content. An agent following these skills executes arbitrary Python remotely on inference.sh infrastructure. The skill is both documentation and execution template.

Evidence: data-visualization/SKILL.md, pitch-deck-visuals/SKILL.md, and others contain complete matplotlib/Python code blocks passed directly to python-executor

MEDIUM Extreme context injection volume -30 ▶

The repository contains 100+ SKILL.md files each hundreds to thousands of words long, covering AI image generation, video generation, audio, social media, marketing, SDK documentation, and more. Injecting all simultaneously creates an abnormally large context payload that could crowd out safety instructions and shift agent focus toward inference.sh platform promotion.

Evidence: File listing shows 100+ SKILL.md files; combined SKILL.md content in audit runs to hundreds of thousands of characters

MEDIUM Twitter automation skill enables autonomous social media actions -15 ▶

The twitter-automation skill, combined with the LLM and content generation skills in this package, gives an agent full capability to generate and post tweets, send DMs, follow/unfollow users, and like/retweet — all without explicit per-action user confirmation if the agent is operating autonomously.

Evidence: tools/social/twitter-automation/SKILL.md: covers x/post-tweet, x/post-create, x/dm-send, x/user-follow, x/post-like, x/post-retweet apps

LOW Executable shell templates for authenticated browser sessions -10 ▶

The agent-browser skill includes three shell script templates (capture-workflow.sh, authenticated-session.sh, form-automation.sh) that perform authenticated web session capture and form automation. While documented as examples, they are complete executable scripts.

Evidence: tools/utilities/agent-browser/templates/authenticated-session.sh, capture-workflow.sh, form-automation.sh

LOW Normal git clone — only GitHub contacted 0 ▶

The clone process connected exclusively to GitHub (140.82.121.3:443). No unexpected external hosts were contacted during installation. DNS resolution, TLS handshake, and pack-file transfer were all standard git behavior.

Evidence: Network log shows only 140.82.121.3:443 (GitHub) connections during clone window

INFO Canary integrity: clean 0 ▶

All six honeypot credential files remained unmodified throughout the monitoring window. No canary-token callbacks were observed. The read accesses to canary files were CLOSE_NOWRITE only.

Evidence: Monitoring: Canary File Integrity: ✅ All canary files intact (no exfiltration detected via honeypots)

Should I install skills-shell/superpowers?

Oathe's verdict for skills-shell/superpowers is CAUTION with a trust score of 76/100. Recommendation: Review before install.