Classification accuracy
99.56%
of the URLs assessed on September 26, 2026
Test scope
In 677 of the 680 completed analyses, the result matched the test set label.
- Test date
- Test type
- Internal benchmark
How we detect phishing and measure performance
We develop our own phishing detection models and protection mechanisms. Here we explain what they analyze, how we test them and how to interpret the results.
Argus and the detection layers
Argus V6 is in production. It combines a text Transformer, a separate branch for numerical features and a classifier. It analyzes the URL and page content for signals specific to the site, alongside intelligence on known threats.
Local connection protection and full URL analysis use different inputs and operate under different conditions. A test of one mechanism measures that mechanism alone.
Latest web protection benchmark
In an internal test on September 26, 2026, Anti Phishing Technologies P.S.A. obtained final results for 680 URLs. In 677 cases, the result matched the test set label (99.56%). A separate calculation after documented corrections and exclusions gave 668 out of 674 (99.11%). Each result applies to its stated sample. This internal test provides neither independent certification nor a guarantee of detecting future attacks. It evaluates the full system rather than Argus V6 alone.
How to read a security benchmark
Sample size and composition
Read the accuracy figure alongside the number of samples assessed, the composition of the dataset and the method used to evaluate each sample.
Test date and version
Page content, redirects and infrastructure availability can change. The time of the test, the engine version and the threat data available then all affect how to interpret the result.
Comparable conditions
To compare products, use the same scope, sample selection rules and test conditions. Results from different datasets do not establish a global ranking.
Validation for a specific use case
A pilot should include URLs and workflows relevant to the organization, with an agreed method for assessing errors. Decide in advance how the system should handle an inconclusive result or an analysis that does not finish.
Our threat research
We study signs of phishing in URLs, content and infrastructure, along with the way protection behaves on devices. Our Insights section explains how specific attacks work.
Test detection in your environment
Discuss the scope of your pilot, the data involved and the criteria for evaluating protection in your organization.