Know if a link is dangerous before anyone clicks.
Agentic cybersecurity for an increasingly insecure world. One API call, and our AI agents tell you whether a URL is safe — and how that determination was made. The result comes back as machine-readable data, so your own agents can weigh it before they act.
1,000 free scans/month · No credit card
curl -X POST http://127.0.0.1:8080/scan
-H "X-API-Key: 01234<api_key>56789"
-H "Content-Type: application/json"
-d '{"url": "http://evil.co/1"}'
{
"url": "http://evil.co/1",
"probability": 0.987787,
"malicious_percent": 98.78,
"is_malicious": true,
"threshold": 0.6,
"source": "model"
}84.47%
Accuracy
92.09%
ROC-AUC
67.81%
TPE
10,000
URLs validated
Why Smarty-Pants
Blocklists are reactive. Our agents think.
Bespoke decision model
A proprietary inference core weighs hundreds of latent signals per link and returns a calibrated posterior on malicious intent — not a list match.
Sub-50ms verdicts
Most URLs are classified in about 20ms, and every verdict lands inside 50ms; deep scans stream results as the agents work.
Explainable, machine-readable
Every verdict ships with findings your analysts can act on and your own agents can read and use directly.
Zero-day coverage
Catch brand-new phishing kits and malicious domains hours before feeds list them.
Simple REST API
One endpoint, clean JSON, callable from any language that speaks HTTP.
Built for scale
Batch scanning, webhooks and 99.99% uptime for high-volume pipelines.
How it works
From URL to verdict in three steps
01
Send a URL
POST any link to /scan or up to 100 at once to /batch from your app, email gateway or SOC pipeline.
02
Agents investigate
A bespoke model evaluates the URL across hundreds of latent signals — content, redirects, certificates, behavior — and makes the decision itself.
03
Act on the verdict
Get a risk score, categories and plain-English evidence to block, warn or allow.
Model performance
Current model: v3.2.0
Validation protocol
How we grade the model
Operational grading is determined on a purposefully curated validation set of 10,000 URLs drawn from known-malicious and benign real-world corpora. The set is withheld from all training and tuning data, preserves the 90/10 legitimate-to-malicious base rate encountered in production traffic, and is scored at a fixed decision threshold. Every figure below is therefore an out-of-sample estimate of deployed behaviour rather than an in-sample measure of fit.
Tuning metric
TPE — True Positive Excess
Accuracy and ROC-AUC characterise discrimination, but neither prices the trade-off we actually operate under. We tune on a custom metric, True Positive Excess: the true-positive rate held in excess of the false-positive rate, penalised by a cost weight α (TPE = TPR − α·FPR). At α = 1.00 one false alarm is charged against exactly one missed malicious URL, so maximising TPE selects the operating point with the highest net detection yield.
Read the TPE paper →0.8447
Accuracy
Overall agreement across 10,000 URLs
0.9209
ROC-AUC
Ranking quality across all thresholds
0.6781
TPE
TPR = 0.8320 · FPR = 0.1539 · α = 1.00
Confusion matrix — 10,000 URLs
7,615
True negative
1,385
False positive
168
False negative
832
True positive
Classification report
| Class | Precision | Recall | F1 | Support |
|---|---|---|---|---|
| legitimate | 0.98 | 0.85 | 0.91 | 9000 |
| malicious | 0.38 | 0.83 | 0.52 | 1000 |
| accuracy | 0.84 | 10000 | ||
| macro avg | 0.68 | 0.84 | 0.71 | 10000 |
| weighted avg | 0.92 | 0.84 | 0.87 | 10000 |
Malicious precision of 0.38 is the arithmetic consequence of a 10% base rate: at a 15.4% false-positive rate, 1,385 legitimate URLs are flagged for every 832 caught. Tuning on TPE rather than precision is what keeps the malicious recall at 0.83.
Testimonials
Teams that stopped worrying about links
"It runs agentically end to end — the agents investigate the link, explain the call, and our gateway acts on it. Phishing reports dropped by 80%."
"The agents do the investigation and show their work, so our analysts trust a verdict instead of re-running it."
"Two endpoints, clear JSON, honest pricing. Exactly what a developer wants."
Our approach
Built by humans & AI to enhance the security of AI
Security researchers and autonomous agents work side by side — so as AI agents browse, buy and act on the web, they don't walk into traps.
Start scanning URLs in minutes
1,000 free checks every month. No credit card required.
Get your API key