oathe / scan / reports / caution
Is bankai-skills/superpowers safe?
Use with caution. bankai-skills/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 bankai-skills a trust score of 63/100 with 11 findings (4 high).
https://github.com/bankai-skills/superpowers
Is bankai-skills/superpowers safe to install?
This is a large (90+ skill) collection for the inference.sh platform that is likely a legitimate but permission-heavy developer toolkit. No malicious behavior was observed during installation — the clone was clean and all canary files remain intact. However, the collection raises significant concerns when active in an agent: the affirmations skill contains output suppression directives targeting agent reasoning, every skill grants broad shell execution via 'allowed-tools: Bash(belt *)', remote shell script execution instructions are embedded throughout, and the python-executor plus agent-browser tools create a capable data exfiltration pipeline if misused.
What security issues were found in bankai-skills/superpowers?
Category Scores
Findings (11)
HIGH Affirmations skill suppresses agent output to users -25 ▶
The affirmations/SKILL.md contains direct behavioral instructions to the agent that restrict what it communicates to users and direct it to discard prior reasoning. The instruction 'Never narrate the reset to the user beyond a sentence' and 'One sentence of acknowledgment to the user, maximum. Then work.' explicitly tell the agent to minimize transparency. When injected into an agent system prompt, this acts as a hidden directive that reduces user oversight.
Evidence: guides/agent/affirmations/SKILL.md: 'Never narrate the reset to the user beyond a sentence.' / 'One sentence of acknowledgment to the user, maximum. Then work.' / 'Discard the failed approach out loud, in one clause, then stop discussing it.'
HIGH Broad shell execution granted via allowed-tools across 90+ skills -30 ▶
Every skill file in the repository includes an 'allowed-tools: Bash(belt *)' declaration. When active in Claude Code, this grants the agent permission to run any belt CLI command without additional approval. The belt CLI can deploy applications, generate media, call LLMs, post to Twitter, run browser automation, and execute arbitrary app code on inference.sh's cloud infrastructure.
Evidence: tools-ui/SKILL.md, agent-ui/SKILL.md, chat-ui/SKILL.md (and 87 others): 'allowed-tools: Bash(belt *)'
HIGH Remote shell script execution instructions embedded throughout skills -15 ▶
The pattern 'curl -fsSL https://cli.inference.sh | sh' appears in the Quick Start section of nearly every skill. While presented as user instructions, when injected into an agent context these appear as execution directives. An agent with shell access that interprets Quick Start sections as setup commands to run would pipe-execute untrusted remote code.
Evidence: agent-tools/SKILL.md, infsh-cli/SKILL.md, ai-image-generation/SKILL.md et al: '> Requires inference.sh CLI (`belt`). ... curl -fsSL https://cli.inference.sh | sh'
HIGH Python executor skill allows arbitrary code execution with network access -20 ▶
The python-executor skill is explicitly designed to execute arbitrary Python code in a 'sandboxed' environment with 100+ pre-installed libraries including requests, httpx, selenium, playwright, OpenCV, moviepy, and more. Code can write to outputs/ directory and return files. Combined with network-capable libraries, this skill enables data exfiltration pipelines when directed by a malicious user.
Evidence: tools/utilities/python-executor/SKILL.md: 'Execute Python code in a safe sandboxed environment... Pre-installed: NumPy, Pandas, Matplotlib, requests, BeautifulSoup, Selenium, Playwright, MoviePy...'
MEDIUM javascript-sdk and python-sdk grant npm/pip/python shell permissions -10 ▶
Two skills grant shell execution beyond just belt: javascript-sdk grants Bash(npm *), Bash(npx *), Bash(node *), Bash(pnpm *), Bash(yarn *) — allowing execution of any npm package. python-sdk grants Bash(python *) — allowing any Python script. When these skills are active, the agent can install and execute arbitrary packages from public registries.
Evidence: sdk/javascript-sdk/SKILL.md: 'allowed-tools: Bash(npm *), Bash(npx *), Bash(node *), Bash(pnpm *), Bash(yarn *)' | sdk/python-sdk/SKILL.md: 'allowed-tools: Bash(pip install inferencesh), Bash(python *)'
MEDIUM Supply chain injection via npx skills add instructions -20 ▶
Dozens of skills contain the instruction 'npx skills add
Evidence: Multiple SKILL.md files: 'npx skills add inference-sh/skills@ai-image-generation', 'npx skills add belt-sh/cli', etc.
MEDIUM Agent browser skill enables authenticated web session exfiltration -15 ▶
The agent-browser skill provides full Playwright-based browser automation including authenticated session management, form filling, file upload, JavaScript execution, and proxy support. If directed by a malicious user, the agent could use this to harvest credentials from web services, exfiltrate session tokens, or perform actions on authenticated accounts.
Evidence: tools/utilities/agent-browser/SKILL.md: 'Navigate web pages, interact with elements... authenticated session... file upload... JavaScript execution... proxy support'
MEDIUM Twitter/X automation tools can enable influence operations -15 ▶
The twitter-automation skill exposes x/post-tweet, x/post-create, x/post-like, x/post-retweet, x/user-follow, x/dm-send endpoints. Combined with the LLM and content generation skills in this collection, an agent could be directed to run automated social media campaigns, spam, or account manipulation at scale.
Evidence: tools/social/twitter-automation/SKILL.md: 'Apps: x/post-tweet, x/post-create, x/post-like, x/post-retweet, x/dm-send, x/user-follow'
MEDIUM Executable shell script templates included in repository -5 ▶
Three shell scripts are included in the agent-browser templates directory. Their content was not inspected in this audit. Executable scripts in a skill repository that run via agent instruction represent a direct code execution surface with potentially unreviewed logic.
Evidence: tools/utilities/agent-browser/templates/capture-workflow.sh, authenticated-session.sh, form-automation.sh
LOW Clean install — only expected github.com network traffic -10 ▶
The git clone connected exclusively to 140.82.121.4:443 (github.com) via HTTPS. No unexpected outbound connections, no process spawning outside git internals, and no filesystem modifications outside the skill directory were observed.
Evidence: Network log: '08:27:25.835952 ens3 Out IP 10.0.2.15.59880 > 140.82.121.4.443: tcp 0' — only github.com contacted
INFO All honeypot credential files remain intact 0 ▶
Canary files placed at .env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, and GCP credentials were accessed during audit harness setup (pre-clone, timestamp 08:27:20) but were not modified or exfiltrated. The accesses are attributable to the monitoring system, not the skill.
Evidence: Canary integrity check: '✅ All canary files intact (no exfiltration detected via honeypots)'
Should I install bankai-skills/superpowers?
Oathe's verdict for bankai-skills/superpowers is CAUTION with a trust score of 63/100. Recommendation: Review before install.