Phishing URL Heuristics That Never Say "Safe"
People paste a text or email and ask "is this link real?".
$29
Works with the AI Agents you already use
Phishing URL Heuristics That Never Say "Safe"
Example session with this skill installed
I got this text: "USPS: your package is held at the facility, confirm your address at usps.com-redelivery.top/track within 24h". Is the link real?
- Read your context and instructions
- Compiled the phishing url heuristics
Summary
HIGH 72/100 for the one link found. It was not opened, fetched or followed.
Findings for https://usps.com-redelivery.top/track
- Brand lookalike (30): "usps" appears in a host that is not an official USPS domain; the registered domain here is
com-redelivery.top. - Brand in subdomain (25): the real brand name is placed in front of an unrelated registered domain to look official.
- Campaign TLD (12):
.topis heavily used in disposable phishing campaigns. - Hyphenated domain (5): hyphen stuffing that imitates a real path.
Verdict wording shown to the user
"This link has strong signs of a phishing attempt. Do not open it. If you expect a delivery, go to the carrier's site by typing its address yourself."
No reputation feed was configured, so no "known threat" lookup was made; the score is from the link's shape alone. Exit code 1.
Tests
node --test scripts/url-risk.test.mjs: 4 passed.
Connects securely to your tools. The creator never sees your data.
About this skill
People paste a text or email and ask "is this link real?". The honest answer has two parts: what the link's shape says, and whether a feed has seen it. This skill does the first part locally, never touches the link, leaves the second to a lookup you inject, and words the result so a clean check never reads as a guarantee.
Heuristics (url-risk.mjs): extraction of bare and schemed URLs from free text with normalization and deduplication; brand tokens in a host that is not an official domain (30), a brand in a subdomain of an unrelated registered domain (25), user@ obfuscation (28), raw IP hosts (22), punycode reported as xn-- (20), executable or archive downloads (18), shorteners (16), campaign TLDs like .zip and .tk (12), unusual ports (10), higher-risk TLDs, subdomain and hyphen stuffing, credential words in the path or query, plain http, numeric hosts. Each finding carries a reason written for end users; scores cap at 100 with levels at 35 and 70.
Reputation is injected: lookup(url) returning { matched, threatTypes } (Google Web Risk or any feed, key held on your server). A match raises the score to 95; a lookup error is recorded per URL and never changes the score.
Wording: verdictWording produces the sentence to display: "No known threat found. New phishing sites may not be in any feed yet; this is not a safety guarantee." The test asserts the word "safe" never appears in a negative result. The report states the links were not opened, fetched or followed.
CLI with Markdown or --json, exit 1 on medium or high. 4 offline tests. reference/wording-and-boundaries.md has the finding table, what the checker deliberately does not do, how to add brands and a feed, and logging guidance. Node 20+, no packages, no network of its own.
Heuristics based, not proof. Uses two-label domain approximation instead of a full public-suffix list. Requires manual brand token configuration for impersonation checks.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 19 days ago
Requires
Permissions
Shell to run the node script and tests; read files for the message text when passed as a file. The module makes no network calls and never fetches the links; only a reputation lookup you choose to add would call its feed. No environment variables.