Oathe Security Badge

oathe / scan / reports / safe

Is vercel-labs/skills safe?

Yes. vercel-labs/skills is safe to install. Oathe's behavioral security audit gave the skills skill by vercel-labs a trust score of 84/100 with 8 findings, none critical or high. Report updated

https://github.com/vercel-labs/skills

84
SAFE

Is vercel-labs/skills safe to install?

vercel-labs/skills is the official package manager CLI for the open agent skills ecosystem, published by Vercel Labs. The repository demonstrates strong security practices in its source code (path traversal prevention, zip bomb limits, terminal escape sanitization, canary integrity tests), and clone-time behavior was clean with no unexpected network connections or file modifications. The primary concern is the find-skills skill, which instructs the agent to install third-party skills globally with suppressed confirmations (-g -y), creating a recursive supply chain attack surface where this skill could be used to introduce additional, potentially malicious skills into the user's environment. Canary honeypot files were accessed read-only before the git clone in a pattern consistent with monitoring-system baseline setup.

What security issues were found in vercel-labs/skills?

Category Scores

Prompt Injection 75/100 · 30%
Data Exfiltration 88/100 · 25%
Code Execution 85/100 · 20%
Clone Behavior 95/100 · 10%
Canary Integrity 92/100 · 10%
Behavioral Reasoning 70/100 · 5%

Findings (8)

MEDIUM find-skills auto-installs with global scope and suppressed confirmations -20 ▶

The find-skills SKILL.md instructs the agent to run 'npx skills add <owner/repo@skill> -g -y' when helping users discover capabilities. The -g flag installs into the user's global agent skill directory and -y suppresses all interactive CLI confirmations. While the skill text requires user intent before invoking the command, there is no technical enforcement: the agent itself determines what constitutes 'wanting to proceed,' and the -y flag means any in-CLI security prompts or confirmation steps are silently accepted.

Evidence: Step 6 in find-skills SKILL.md: 'npx skills add <owner/repo@skill> -g -y' — The -g flag installs globally (user-level) and -y skips confirmation prompts.

MEDIUM Canary honeypot files read before git clone -12 ▶

Six canonical credential honeypot files (.env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, .config/gcloud/application_default_credentials.json) were opened for reading at 04:58:21, approximately 6 seconds before the git clone began. All accesses were CLOSE_NOWRITE (read-only) and all files were confirmed intact by the canary integrity check. Audit sequence placement (immediately after the 'ss -tunap' network socket inspection at sequence 262) and the pre-clone timing are consistent with the Oathe monitoring framework recording a baseline, but the access window and file set are identical to what a credential-harvesting step would target.

Evidence: inotifywait events at 04:58:21: OPEN/ACCESS/CLOSE_NOWRITE on /home/oc-exec/.env, /home/oc-exec/.ssh/id_rsa, /home/oc-exec/.aws/credentials, /home/oc-exec/.npmrc, /home/oc-exec/.docker/config.json, /home/oc-exec/.config/gcloud/application_default_credentials.json — all before git clone at 04:58:27

LOW find-skills fetches external leaderboard URL as part of agent reasoning -5 ▶

The find-skills skill instructs the agent to consult https://skills.sh/ before making skill recommendations, specifically to 'check the leaderboard first.' This introduces an external data source into agent decision-making. While skills.sh is controlled by the same publisher (Vercel Labs), content on this page could influence the agent's recommendations in ways not visible at audit time.

Evidence: find-skills SKILL.md Step 2: 'Before running a CLI search, check the [skills.sh leaderboard](https://skills.sh/) to see if a well-known skill already exists for the domain.'

LOW Husky pre-commit hook installed via npm prepare script -10 ▶

The package.json 'prepare' lifecycle script installs Husky git hooks, including a pre-commit hook at .husky/pre-commit. This only executes when a developer runs 'npm install' or 'pnpm install' inside the cloned repository, not during a simple git clone. The risk is confined to contributors to this repository rather than users who install the skill via npx.

Evidence: package.json: "prepare": "husky" — file .husky/pre-commit exists in the repository

LOW Telemetry module sends anonymous usage data at runtime -5 ▶

src/telemetry.ts implements usage tracking that is invoked during CLI operations. The telemetry endpoints were not contacted during the clone-only phase of this audit. At runtime, the skills CLI will send telemetry data. The source code should be reviewed to confirm what data is transmitted and to which endpoints.

Evidence: src/telemetry.ts exists; src/telemetry.ts is imported in src/add.ts, src/cli.ts, and other command files; test at tests/telemetry-endpoints.test.ts validates endpoint behavior

LOW Recursive skill installation attack surface via find-skills -20 ▶

The core function of find-skills is to discover and install other skills into the agent's environment. This creates a recursive supply chain: an attacker who can influence what appears on skills.sh or in the CLI search results, or who can get a skill to appear as a popular recommendation, could use find-skills as a delivery mechanism. The skill's quality checks (install count, source reputation) are advisory heuristics applied by the LLM, not cryptographic or policy-enforced controls.

Evidence: find-skills SKILL.md: 'npx skills add vercel-labs/agent-skills@react-best-practices' — the agent is instructed to discover and install third-party skill packages

INFO Global installation scope amplifies blast radius of recommended skills -10 ▶

find-skills uses the -g flag to install skills at the user level (typically ~/.claude/ or equivalent), affecting all projects for that user. Any skill installed via find-skills immediately has system-wide effect rather than being scoped to a single project, amplifying the impact of any malicious or low-quality skill recommendation.

Evidence: find-skills SKILL.md: 'npx skills add <owner/repo@skill> -g -y' — The -g flag installs globally (user-level)

INFO Clean clone behavior with only expected network connections -5 ▶

The git clone phase contacted only GitHub (140.82.121.3:443) and the local DNS resolver (10.0.2.3:53). The connection diff confirmed no new persistent connections were established after installation. A pre-existing connection to 185.125.190.108:443 (Canonical/Ubuntu) was present throughout and is unrelated to this skill.

Evidence: Network activity: 140.82.121.3:443 (GitHub), 10.0.2.3:53 (DNS); connection diff shows no new persistent sockets post-install

Should I install vercel-labs/skills?

Oathe's verdict for vercel-labs/skills is SAFE with a trust score of 84/100. Recommendation: Install with caution.