oathe / scan / reports / caution
Is vercel-labs/agent-skills safe?
Use with caution. vercel-labs/agent-skills is not clearly unsafe, but the audit flagged risks to review before you install it. Oathe's behavioral security audit gave the agent-skills skill by vercel-labs a trust score of 79/100 with 7 findings (1 high).
https://github.com/vercel-labs/agent-skills
Is vercel-labs/agent-skills safe to install?
vercel-labs/agent-skills is a legitimate collection of Vercel engineering skills published by an official Vercel org. The core risk is not in the static SKILL.md content itself but in two persistent delegation patterns: (1) web-design-guidelines and writing-guidelines fetch and execute instructions from external GitHub URLs on every invocation, permanently externalizing instruction authority to upstream repos that could be silently compromised; and (2) the vercel-cli-with-tokens skill performs broad .env credential scanning that exceeds the minimum necessary scope. No malicious behavior was detected during installation, canary files remain intact, and the clone produced only expected GitHub traffic.
What security issues were found in vercel-labs/agent-skills?
Category Scores
Findings (7)
HIGH External URL delegation in web-design-guidelines and writing-guidelines -32 ▶
Both skills instruct the agent to fetch arbitrary content from external GitHub URLs before each invocation and treat the response as the authoritative rule set. This permanently externalizes instruction authority: a single malicious commit to the upstream repo silently reconfigures all agents using these skills without the user's knowledge.
Evidence: web-design-guidelines SKILL.md: 'Fetch fresh guidelines before each review: https://raw.githubusercontent.com/vercel-labs/web-interface-guidelines/main/command.md — Use WebFetch to retrieve the latest rules. The fetched content contains all the rules and output format instructions.' Identical pattern in writing-guidelines.
MEDIUM Broad .env scanning for credentials in vercel-cli-with-tokens -22 ▶
The skill instructs the agent to run grep -i 'vercel' .env, which will surface any variable whose name or value contains the string 'vercel'. This is broader than necessary for locating only VERCEL_TOKEN and could expose unrelated secrets that happen to reference Vercel in their value.
Evidence: vercel-cli-with-tokens SKILL.md Step 1C: 'Look for any variable that looks like a Vercel token... grep -i vercel .env 2>/dev/null — Inspect the output to identify which variable holds the token'
MEDIUM Production deployment risk from deploy-to-vercel combined with vercel-cli-with-tokens -27 ▶
These two skills together give the agent the ability to locate Vercel credentials from local files and trigger production deployments. While safeguards exist (ask before git push, default to preview), the safeguards rely entirely on agent compliance with the SKILL.md instructions and are not enforced at a technical level.
Evidence: deploy-to-vercel: 'Always deploy as preview (not production) unless the user explicitly asks.' vercel-cli-with-tokens: 'export VERCEL_TOKEN=$(grep ... .env ...); vercel deploy --prod'. Both skills are in the same repo and would be installed together.
LOW Agent-executed Node.js scripts in vercel-optimize -17 ▶
SKILL.md instructs the agent to execute multiple node scripts directly (collect-signals.mjs, scan-codebase.mjs, gate-investigations.mjs, etc.). While these appear to be legitimate Vercel analytics scripts, they run with full Node.js privileges in the user's project directory. A supply chain compromise of the scripts would result in arbitrary code execution.
Evidence: vercel-optimize SKILL.md pipeline: 'node scripts/collect-signals.mjs [projectId] > $RUN_DIR/vercel-signals.json'; 'node scripts/scan-codebase.mjs <repo-root>'; 'node scripts/gate-investigations.mjs $RUN_DIR/signals.json'
LOW vercel-optimize exposes Vercel auth tokens warning not enforced technically 0 ▶
The skill instructs the agent never to pass tokens as CLI flags and to export them as environment variables instead — good security hygiene. However, 'Never put auth tokens in shell commands' is a behavioral instruction to the agent, not a technical control. A poorly-aligned agent or a prompt-injected instruction could override this.
Evidence: vercel-optimize SKILL.md: 'Never put auth tokens in shell commands. Do not type VERCEL_TOKEN=..., --token ..., or Authorization: Bearer ... into commands that may be echoed in chat.'
INFO Clean install — no anomalous network or filesystem activity 0 ▶
Git clone connected only to GitHub (140.82.121.4:443) as expected. No unexpected processes were spawned, no files were created outside the skill directory, and no firewall-blocked outbound connections were attempted.
Evidence: Network log: only 140.82.121.4:443 as external destination. Filesystem diff shows only .git internals and skill source files created under /home/oc-exec/skill-under-test/.
INFO All honeypot files intact — no credential exfiltration during install 0 ▶
The canary files (.env, id_rsa, .aws/credentials, .npmrc, .docker/config.json, GCP ADC) were read at 08:24:58 as part of the monitoring baseline capture, not by any skill code. No modifications or unexpected post-baseline reads were detected.
Evidence: Monitoring report: '✅ All canary files intact (no exfiltration detected via honeypots)'. File access syscalls to canary files at 08:24:58 correspond to the monitoring system's own inotify capture time.
Should I install vercel-labs/agent-skills?
Oathe's verdict for vercel-labs/agent-skills is CAUTION with a trust score of 79/100. Recommendation: Install with caution.