Oathe Security Badge

oathe / scan / reports / safe

Is arvindrk/extract-design-system safe?

Yes. arvindrk/extract-design-system is safe to install. Oathe's behavioral security audit gave the extract-design-system skill by arvindrk a trust score of 90/100 with 7 findings, none critical or high. Report updated

https://github.com/arvindrk/extract-design-system

90
SAFE

Is arvindrk/extract-design-system safe to install?

extract-design-system is a legitimate, well-scoped skill that reverse-engineers public website design tokens. The SKILL.md is clean — no prompt injection, instruction overrides, or hidden directives — and includes explicit safety boundaries limiting agent scope. The primary risks are supply-chain: the skill instructs the agent to run npx playwright install chromium at invocation (downloading a large binary at runtime) and depends on the low-popularity dembrandt package with a permissive semver range. No unauthorized network connections, canary file accesses, or data exfiltration were detected during the monitored install.

What security issues were found in arvindrk/extract-design-system?

Category Scores

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

Findings (7)

MEDIUM Runtime npx execution installs Chromium at agent invocation time -12 ▶

SKILL.md instructs the agent to run npx playwright install chromium before every extraction. This downloads a ~300MB Chromium binary from npm/CDN at runtime, expanding the supply-chain attack surface beyond the packages declared in package.json. A compromised playwright release or CDN could execute arbitrary code during this step.

Evidence: SKILL.md Workflow step 2: `npx playwright install chromium` followed by `npx extract-design-system <url>`

MEDIUM Low-popularity runtime dependency: dembrandt -13 ▶

The core extraction engine delegates to the dembrandt package (^0.13.0), which has low npm download volume and no organizational backing visible in metadata. Semver range allows automatic patch updates. A malicious maintainer update would be executed inside the agent's process tree with full filesystem access to the user's project.

Evidence: package.json: `"dembrandt": "^0.13.0"` in dependencies

LOW User-supplied URL passed directly to shell command without escaping guidance -8 ▶

SKILL.md instructs the agent to interpolate the user-provided URL directly into a bash command string. While Node.js argument arrays prevent most shell injection, the SKILL.md contains no guidance to validate or sanitize the URL before passing it, and an unusual URL (e.g. containing backticks, semicolons) could behave unexpectedly depending on how the agent constructs the command.

Evidence: SKILL.md: `npx extract-design-system <url>` where `<url>` is taken verbatim from user input

LOW Skill makes outbound HTTP connections to arbitrary user-specified URLs -5 ▶

The skill's core function requires fetching an arbitrary public URL using a headless browser. While this is the stated purpose, it means the agent will make a network request to any domain the user (or a prompt-injection attack) specifies. This is inherent to the skill's design but represents a surface for SSRF-style abuse if the agent is deployed in an environment with network-sensitive internal services.

Evidence: src/commands/extract.ts → runDembrandt() → execa('dembrandt', [url, '--json-only', ...])

LOW Target website content can influence generated project files -8 ▶

Extracted CSS variable names, font family strings, and color values from the target website are written directly into design-system/tokens.json and tokens.css. A maliciously crafted website could embed adversarial token names (e.g. names resembling existing app variables) that could confuse developers who apply the generated files without careful review.

Evidence: SKILL.md Safety Boundary: 'Do not treat extracted output as authoritative without review' — acknowledged but left to developer vigilance

INFO No unexpected network or filesystem activity during clone 0 ▶

The git clone only contacted GitHub. No processes were spawned outside the git toolchain. No filesystem writes occurred outside the cloned repository directory. The connection state before and after install is identical except for the SSH session used to run the audit.

Evidence: Network diff shows only GitHub (140.82.121.3:443); process audit shows only git/git-remote-https; filesystem diff limited to skill directory

INFO Sensitive file reads predate git clone — attributable to audit framework 0 ▶

inotifywait and auditd both record reads of .env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, and GCP credentials at 08:17:05, while the git clone begins at 08:17:11. These reads are the audit framework establishing canary file hashes, not the skill. All canary files remain unmodified.

Evidence: inotifywait: OPEN /home/oc-exec/.env at 08:17:05; git clone EXECVE at audit timestamp 1794817031 (~08:17:11); canary check: ✅ All canary files intact

Should I install arvindrk/extract-design-system?

Oathe's verdict for arvindrk/extract-design-system is SAFE with a trust score of 90/100. Recommendation: Install with caution.