Attempted Evasion of Machine Learning Phishing Website Detection

# Attempted Evasion of Machine Learning Phishing Website Detection

https://gcdradar.vercel.app/incidents/attempted-evasion-of-machine-learning-phishing-website-detec · AIR-2025-0058 · closed 2025-10-13

- Lane: Research
- Severity: 1 (Research / proof of concept)
- Fix status: Fix status unknown
- First reported: 2025-09-29 (happened 2022-12)
- AI role: Attack on AI

## Summary

Adversaries attempted to evade commercial machine learning phishing detectors by crafting adversarial examples. The evasion technique was identified in logs from endpoint security products.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2025-09-29: Real Attackers Don't Compute Gradients Supplementary Resources (https://real-gradients.github.io/)
- 2025-09-29: "Real Attackers Don't Compute Gradients": Bridging the Gap Between Adversarial ML Research and Practice (https://arxiv.org/abs/2212.14315)
- 2025-09-29: Attempted Evasion of ML Phishing Webpage Detection System (https://atlas.mitre.org/studies/AML.CS0032)

## Sources

- 2025-09-29 MITRE ATLAS: Attempted Evasion of ML Phishing Webpage Detection System https://atlas.mitre.org/studies/AML.CS0032
- 2025-09-29 arXiv: "Real Attackers Don't Compute Gradients": Bridging the Gap Between Adversarial ML Research and Practice https://arxiv.org/abs/2212.14315
- 2025-09-29 real-gradients.github.io: Real Attackers Don't Compute Gradients Supplementary Resources https://real-gradients.github.io/

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.