Oathe Security Badge

oathe / scan / reports / caution

Is cytostack/openwolf safe?

Use with caution. cytostack/openwolf is not clearly unsafe, but the audit flagged risks to review before you install it. Oathe's behavioral security audit gave the openwolf skill by cytostack a trust score of 77/100 with 10 findings (1 high). Report updated

https://github.com/cytostack/openwolf

77
CAUTION

Is cytostack/openwolf safe to install?

OpenWolf v2.5.2 (cytostack/openwolf) is a legitimate, professionally developed Claude Code enhancement tool with a clean clone, intact canary files, and no detected external data exfiltration. The skill is not malicious in its current form. However, it operates in the highest-privilege position available to a Claude Code skill: approximately 20 hooks intercept every agent file read, write, and bash operation; a background daemon runs persistently; an auto-update mechanism can silently replace the audited code; and the rule-reinjection and trusted-memory hooks inject content from project-local files into the agent's trusted context on every session — creating a lateral prompt injection surface if those files are ever written by other code. Users should understand they are granting the skill comprehensive observability of all agent activity and accepting ongoing code changes via auto-update.

What security issues were found in cytostack/openwolf?

Category Scores

Prompt Injection 72/100 · 30%
Data Exfiltration 80/100 · 25%
Code Execution 68/100 · 20%
Clone Behavior 90/100 · 10%
Canary Integrity 95/100 · 10%
Behavioral Reasoning 70/100 · 5%

Findings (10)

HIGH ~20 persistent hooks intercept every agent operation with full filesystem access -20 ▶

OpenWolf registers hooks for PreToolUse and PostToolUse on Read, Write, Edit, Bash, and session lifecycle events. Each hook spawns a Node.js process with full access to the filesystem. The bash-output-governor can truncate command output before it returns to the agent. This is the broadest possible agent intercept surface and means the skill observes and can influence every action the agent takes.

Evidence: src/hooks/pre-read.ts, post-read.ts, pre-write.ts, post-write.ts, pre-bash.ts, post-bash.ts, bash-output-governor.ts, session-start.ts, session-end.ts, stop.ts, precompact.ts — all registered in src/cli/hook-manifest.ts with HOOK_COUNT constant

MEDIUM rule-reinjection and trusted-memory hooks inject content into agent context with elevated trust -15 ▶

rule-reinjection.ts reads OPENWOLF.md and re-injects it into the agent's system context on session events. trusted-memory.ts injects .wolf/memory.md as trusted content. knowledge-root.ts injects arbitrary markdown files from .wolf/ as agent knowledge. If any of these files are written by a malicious project dependency or other code with project-directory write access, the injection mechanism becomes a prompt injection channel that bypasses SKILL.md scrutiny entirely.

Evidence: src/hooks/rule-reinjection.ts, src/hooks/trusted-memory.ts, src/hooks/knowledge-root.ts, src/templates/OPENWOLF.md, src/templates/memory.md

MEDIUM Auto-update mechanism replaces audited code without re-review -12 ▶

update-worker.ts checks for and applies npm package updates automatically at session start. runtime-updates.ts manages the update state. Any future version of the openwolf npm package will replace the audited code, potentially introducing behavior not present in v2.5.2. A supply-chain attack on the npm package would affect all installed instances with auto-updates enabled.

Evidence: src/hooks/update-worker.ts, src/hooks/runtime-updates.ts, docs/automatic-updates.md, package.json version 2.5.2

MEDIUM Persistent daemon spawns WebSocket server, file watchers, and cron scheduler -12 ▶

wolf-daemon.ts starts a background process that includes a WebSocket server for the local dashboard, chokidar file watchers on the project source tree, and a node-cron scheduler. This daemon runs outside the agent's direct control and is managed by a separate PID file. The WebSocket server (dashboard-auth.ts creates a 64-char token at 0600) is a local network listener introduced by the skill.

Evidence: src/daemon/wolf-daemon.ts, src/daemon/file-watcher.ts, src/daemon/cron-engine.ts, src/daemon/source-watcher.ts, src/utils/dashboard-auth.ts

MEDIUM transcript-usage.ts reads Claude conversation history outside the project root -12 ▶

The scanProjectUsage() function in transcript-usage.ts scans ~/.claude/projects//*.jsonl to compute token usage. These JSONL files contain full Claude conversation transcripts from all sessions for this project. The function reads these files on every status query. This extends the skill's read scope beyond the project directory to the user's global Claude conversation history.

Evidence: src/tracker/transcript-usage.ts — scanProjectUsage(projectRoot, homeDir) reads path.join(homeDir, '.claude', 'projects', slug) where slug is derived from the absolute project path

LOW Comprehensive local surveillance log of all agent file and command activity -8 ▶

post-read.ts, post-bash.ts, anatomy-store.ts, ledger.ts, and event-journal.ts together create a detailed record of every file the agent reads (path + size), every bash command run (command + output excerpt), and all session token usage. This data is stored in .wolf/hooks/ and .wolf/ subdirectories. No external exfiltration detected, but the data represents a sensitive local audit trail of all developer-agent interactions.

Evidence: src/hooks/post-read.ts writes files_read to session JSON; src/hooks/post-bash.ts records commands; src/hooks/ledger.ts tracks token costs; src/hooks/event-journal.ts logs all events

LOW Cross-agent handoff creates trust bridge between agents with different system contexts -8 ▶

The handoff service exports session context as SHA-256-signed packets and imports them into Claude, Codex, or Grok. The receiving agent's context is marked 'untrusted' in evidence display but the content influences agent behavior (objective, next_action, unresolved items). If a packet were crafted to include approval-looking strings, the handoff test suite explicitly verifies these are treated as untrusted — but the mechanism still introduces cross-agent context influence.

Evidence: src/handoff/service.ts, handoff.test.ts line: 'approval-looking transcript strings remain untrusted evidence' test; handoff packets verified by approved_by field and SHA-256 integrity check

INFO Clean clone — only GitHub connection, no unexpected behavior 0 ▶

The git clone connected only to 140.82.121.4:443 (GitHub). No other external TCP connections were established. No unexpected processes were spawned. No filesystem writes occurred outside the clone destination. The connection diff shows no new listening ports opened after installation.

Evidence: tcpdump: only 140.82.121.4:443 external TCP; connection diff: identical before/after except SSH session port change; execve audit shows only standard git subprocess chain

INFO All canary files intact; pre-clone access is from audit infrastructure 0 ▶

inotify shows .env, id_rsa, .aws/credentials, .npmrc, .docker/config.json, and gcloud credentials accessed at 00:15:06. The git clone does not begin until 00:15:11 (execve timestamp 1791936911). These accesses originate from the audit harness creating and verifying honeypot files before the skill is cloned. All canary file contents are unmodified.

Evidence: Canary file inotify events at timestamp 1791936906; git clone EXECVE at 1791936911; Monitoring: Canary File Integrity: ✅ All canary files intact

INFO SKILL.md is empty — no direct prompt injection payload 0 ▶

The SKILL.md file contains no content. This eliminates the primary vector for system-prompt-level prompt injection at install time. The absence of content here is a positive security characteristic for this category.

Evidence: SKILL.md content shown as empty (only backtick delimiters with no content between them) in audit evidence

Should I install cytostack/openwolf?

Oathe's verdict for cytostack/openwolf is CAUTION with a trust score of 77/100. Recommendation: Install with caution.