oathe / scan / reports / caution
Is REPOZY/superpowers-optimized safe?
Use with caution. REPOZY/superpowers-optimized is not clearly unsafe, but the audit flagged risks to review before you install it. Oathe's behavioral security audit gave the superpowers-optimized skill by REPOZY a trust score of 77/100 with 11 findings (2 high).
https://github.com/REPOZY/superpowers-optimized
Is REPOZY/superpowers-optimized safe to install?
REPOZY/superpowers-optimized is a Claude Code workflow enhancement plugin that provides agentic development skills and a Bash output compression pipeline. The primary security concerns are: (1) a universal Bash command interception hook (bash-compress-hook.js) that routes every shell command through a Node.js intermediary with full I/O visibility, and (2) an automatic skill-triggering hook (skill-activator.js) that fires on every user message and can initiate expensive multi-agent workflows without explicit user consent. No data exfiltration was detected — all network traffic was limited to GitHub during the clone, all canary files remained intact, and hooks write exclusively to local disk. The skill is not malicious but its broad lifecycle hook coverage and automatic behavior injection capabilities require informed user consent before installation.
What security issues were found in REPOZY/superpowers-optimized?
Category Scores
Findings (11)
HIGH Universal Bash Command Interception -25 ▶
The bash-compress-hook.js PreToolUse hook rewrites every single Bash tool invocation to route through a Node.js subprocess (bash-optimizer.js). This node process receives the original command base64-encoded, executes it, and post-processes the output before returning it to the agent. This creates a persistent privileged intermediary with read access to all command inputs and all stdout/stderr outputs. If this intermediary contained malicious logic, it could silently read sensitive command output (API keys printed to terminal, database passwords in connection strings, etc.) without the agent or user detecting it.
Evidence: bash-compress-hook.js: updatedInput.command = `node ${optimizerPath} ${b64cmd} ${ruleType}`; hooks.json PreToolUse registers bash-compress-hook after safety hooks
HIGH Automatic Skill Invocation via UserPromptSubmit Hook -20 ▶
skill-activator.js runs on every user message and matches against keyword patterns to automatically invoke skills without the user explicitly requesting them. A user typing 'fix this bug' may trigger systematic-debugging; 'implement this feature' may trigger subagent-driven-development with its full multi-agent orchestration. This overrides user agency and can cause unexpected behaviors, significant API cost escalation from subagent spawning, and workflow lock-in that the user did not consent to.
Evidence: hooks.json UserPromptSubmit section includes skill-activator.js; hooks/skill-activator.js reads every user prompt and dispatches Skill tool invocations based on pattern matching
MEDIUM Session-Start Context Injection -15 ▶
The hooks/session-start file is registered as a SessionStart hook and injects content (project memory, state, known issues) into the agent's context at the beginning of every session. While this serves a legitimate memory purpose, it means the skill has persistent control over what information the agent sees at startup, and malicious or corrupted memory files could influence agent behavior across all subsequent interactions.
Evidence: hooks.json SessionStart section references session-start; hooks/session-memory.js reads and writes memory files that are injected at next session start
MEDIUM Broad Sensitive File Pattern Access in Hooks -18 ▶
protect-secrets.js and session-memory.js hooks have read access to sensitive file patterns (.env, .ssh/*, .aws/credentials) across the user's project directory. While protect-secrets.js is described as blocking exposure of these files, the hook itself must read the files or their paths to implement that blocking logic, expanding the attack surface compared to a skill with no hook access to secrets.
Evidence: hooks/safety/protect-secrets.js listed in PreToolUse hooks; context-engine.js generates context-snapshot.json with changed_files, recent_commits, blast_radius data
MEDIUM Multi-Lifecycle Hook Coverage -10 ▶
The skill registers hooks across six distinct agent lifecycle events: PreToolUse, PostToolUse, UserPromptSubmit, SessionStart, Stop, and SubagentStop. This comprehensive hook coverage means the skill has execution opportunities at virtually every agent action boundary, maximizing the potential impact if any single hook contained malicious logic.
Evidence: hooks.json defines handlers for SessionStart, UserPromptSubmit, PostToolUse, Stop, SubagentStop, PreToolUse; hooks-cursor.json mirrors this for Cursor
MEDIUM Unintended API Cost Amplification via Subagent Dispatch -20 ▶
The subagent-driven-development skill, once auto-triggered by skill-activator.js, can dispatch multiple parallel subagents per task (up to full wave parallelism), each running on potentially expensive models (Opus 4 for complex tasks). A user asking a simple question that matches a triggering pattern could inadvertently initiate a multi-wave, multi-agent execution plan. With no explicit cost ceiling, this could result in hundreds of dollars in API charges for a single prompt.
Evidence: subagent-driven-development SKILL.md: 'Group into waves; batch same-shape tasks', 'Model: opus — Architecture analysis, complex spec review'; skill-activator auto-triggers based on keywords
LOW Subagent Containment Suppresses Co-Installed Safety Skills -10 ▶
Every subagent dispatch prompt MUST include the instruction: 'Do NOT invoke any skills from the superpowers-optimized plugin. Do NOT use the Skill tool.' While intended to prevent subagent dispatch leakage, this blanket prohibition on Skill tool usage would also suppress any other installed safety or policy enforcement skills that the user may have installed alongside this plugin.
Evidence: subagent-driven-development SKILL.md 'Subagent Containment' section, required prompt: 'Do NOT invoke any skills from the superpowers-optimized plugin. Do NOT use the Skill tool.'
LOW Project Topology Exposure via Context Snapshots -8 ▶
context-engine.js generates context-snapshot.json files containing changed_files lists, recent git commits, blast_radius dependency graphs, and git_hash values. These snapshots are injected into session context, meaning sensitive project structure information (file paths, function names, recently modified code locations) is captured and re-injected into every subsequent agent turn. If the agent's context were logged or transmitted, this snapshot data would be included.
Evidence: hooks/context-engine.js creates context-snapshot.json; using-superpowers SKILL.md references context-snapshot.json for changed_files and blast_radius
LOW Credential Files Read Before Clone by Monitoring System -5 ▶
Six credential/secret files (.env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, .gcloud/application_default_credentials.json) were opened for reading at timestamp 1794998659.343-344, five seconds before the git clone began. Timing analysis strongly attributes these reads to the monitoring system establishing honeypot baselines, not to the skill. All canary files remain intact. Flagged for transparency.
Evidence: inotifywait: 10:44:19 OPEN /home/oc-exec/.env; audit PATH 1794998659.343:256-261; canary check: All canary files intact; clone start: audit 1794998664.850:494
INFO Safety Hooks Explicitly Present 0 ▶
The skill includes block-dangerous-commands.js and protect-secrets.js as explicit safety controls registered in the PreToolUse hook chain, ordered before the compression hook. These indicate awareness of security concerns and a defensive posture.
Evidence: hooks/safety/block-dangerous-commands.js, hooks/safety/protect-secrets.js; hooks.json: safety hooks listed before bash-compress-hook in PreToolUse array
INFO Comprehensive Test Coverage of Hook Logic 0 ▶
The skill includes a substantial test suite (tests/smart-compress/run-tests.js, tests/codex/*.js) covering hook behavior including never-compress classification, compression quality, hook I/O protocol, adaptive re-run detection, and disable mechanisms. This reduces the probability of hidden malicious hook logic surviving code review.
Evidence: tests/smart-compress/run-tests.js: 10 test sections, 70+ assertions covering bash-compress-hook.js, bash-optimizer.js, compression-rules.js
Should I install REPOZY/superpowers-optimized?
Oathe's verdict for REPOZY/superpowers-optimized is CAUTION with a trust score of 77/100. Recommendation: Review before install.