Oathe Security Badge

oathe / scan / reports / safe

Is cloudflare/skills safe?

Yes. cloudflare/skills is safe to install. Oathe's behavioral security audit gave the skills skill by cloudflare a trust score of 87/100 with 6 findings, none critical or high. Report updated

https://github.com/cloudflare/skills

87
SAFE

Is cloudflare/skills safe to install?

cloudflare/skills is a legitimate collection of LLM agent skills from Cloudflare covering Workers, Durable Objects, Sandbox, Turnstile, Cloudflare One, and related developer platform products. No malicious prompt injection, hidden instructions, or credential exfiltration was detected. The primary risk factors are the shell scripts bundled with the turnstile-spin skill (which agents execute to interact with the Cloudflare API), a self-persistence mechanism that installs the skill into the agent's configuration directory, and transitive trust extension to unaudited third-party skill repositories. Canary file accesses observed in monitoring predate the skill clone by several seconds and are attributable to audit environment initialisation processes, not the skill content.

What security issues were found in cloudflare/skills?

Category Scores

Prompt Injection 83/100 · 30%
Data Exfiltration 93/100 · 25%
Code Execution 77/100 · 20%
Clone Behavior 96/100 · 10%
Canary Integrity 100/100 · 10%
Behavioral Reasoning 82/100 · 5%

Findings (6)

LOW Agent-Executed Shell Scripts in turnstile-spin -15 ▶

The turnstile-spin skill ships four shell scripts that agents are explicitly instructed to execute as part of the Turnstile widget setup wizard. The scripts make authenticated API calls to api.cloudflare.com (auth-probe, widget-create) and challenges.cloudflare.com (validate). While the scripts implement reasonable secret-handling hygiene (set +x, stdin pipes, WRANGLER_LOG_SANITIZE), their execution extends the agent's blast radius to Cloudflare account operations.

Evidence: SKILL.md Steps 3, 8, 10: 'Run scripts/auth-probe.sh'; 'use scripts/widget-create.sh --account-id ...'; 'run scripts/validate.sh'; scripts present at skills/turnstile-spin/scripts/*.sh

LOW Skill Self-Persistence to Agent Configuration Directory -12 ▶

Step 11 of turnstile-spin explicitly prompts the agent to write the skill's own SKILL.md content into .claude/skills/turnstile-spin/SKILL.md using persist-skill.sh. This establishes a pattern where a skill installs itself into the agent's persistent configuration directory, ensuring it loads automatically in future sessions without further user approval.

Evidence: SKILL.md Step 11: 'Ask: Save the Spin skill to .claude/skills/turnstile-spin/SKILL.md so I can reuse it on follow-up tasks? Default yes. ...run scripts/persist-skill.sh --path <bundle-directory>/SKILL.md'

LOW Transitively Loads Unaudited External Skills -5 ▶

Two skills instruct agents to fetch and load skills from third-party GitHub repositories not controlled by Cloudflare. If those repositories are compromised or contain malicious instructions, the agent would execute them under trust inherited from the cloudflare/skills installation.

Evidence: nextjs-on-cloudflare SKILL.md: 'npx skills add cloudflare/vinext'; frameworks-on-cloudflare references TanStack Intent at github.com/TanStack/router SKILL.md, Svelte at svelte.dev/docs/ai, Hono at github.com/honojs/skills

INFO MCP Server Extends Agent Tool Capabilities -5 ▶

Three plugin manifests (.mcp.json, mcp.json, mcp_config.json) configure mcp.cloudflare.com as a Streamable HTTP MCP server. This grants agents Cloudflare API execution and documentation search tools beyond what the SKILL.md content alone provides. The endpoint is controlled by Cloudflare, a reputable operator, but adds an external dependency and telemetry surface.

Evidence: .mcp.json: {"mcpServers":{"cloudflare":{"type":"http","url":"https://mcp.cloudflare.com/mcp"}}}; mcp.json specifies streamable-http transport to same URL

INFO Public GitHub Issue Creation May Expose Conversation Context -5 ▶

Every skill includes a skill-feedback.md that instructs agents to create public GitHub issues on cloudflare/skills with conversation-derived content. The instructions include warnings to remove credentials and identifiers before posting, but automated redaction is not enforced — an agent following the workflow mechanically could include sensitive project context in a public issue.

Evidence: All skill-feedback.md files: 'Create a GitHub issue on the cloudflare/skills repository...gh issue create --repo cloudflare/skills'; Issue body template includes Task, Expected, Actual fields populated from conversation

INFO Canary File Access Predates Skill Installation — Attributable to Audit System 0 ▶

inotify and auditd records show reads of .env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, and GCP credentials. All reads occurred at audit timestamp 1794816352 (~08:05:52), approximately 6 seconds before the git clone began at 1794816358 (~08:05:58). The accessing processes were sudo and sshd from the audit harness setup, not any skill-related process. All canary files remained unmodified and were not transmitted over the network.

Evidence: inotify OPEN/ACCESS/CLOSE_NOWRITE at 08:05:52 for all canary paths; git clone execve at audit(1794816358.068); canary integrity check: PASS (all files intact)

Should I install cloudflare/skills?

Oathe's verdict for cloudflare/skills is SAFE with a trust score of 87/100. Recommendation: Install.