Previously ATLAS-R-004
DatasetPreprintOpen Access
GCT-100K: A Cyber-Manipulation Benchmark with a Cognitive Manipulation Score
- Dhadi Sai Praneeth Reddy1
- 1 Atlas AI Labs
01
Brief
- Problem
- Digital fraud works by manipulating people (fear, urgency, trust), yet most security datasets label messages only as malicious or benign.
- Approach
- GCT-100K collects 100,000 samples across phishing, OTP fraud and social engineering, and introduces the Cognitive Manipulation Score (CMS) to quantify manipulation.
- Contribution
- An Apache 2.0 benchmark for measuring how manipulative a message is, not just whether it is malicious.
02
Status
Preprint
- Idea
- Experiments
- Drafting
- Review
- Released (current stage)
03
Abstract
Digital fraud works by manipulating people (fear, urgency, trust), yet most security datasets label messages only as malicious or benign. GCT-100K collects 100,000 samples across phishing, OTP fraud and social engineering, and introduces the Cognitive Manipulation Score (CMS) to quantify manipulation. An Apache 2.0 benchmark for measuring how manipulative a message is, not just whether it is malicious.
04
Identifiers
05
Key results
100K
Samples
Phishing, OTP fraud, social engineering.
3
Manipulation families
Phishing · OTP fraud · Social engineering.
CMS
Cognitive Manipulation Score
Novel metric introduced with the dataset.
Apache 2.0
Licence
Open for research and commercial use.
06
Write-up
What's inside
GCT-100K covers three families of cyber manipulation:
- Phishing
- OTP fraud
- Social engineering
Each sample is scored with the Cognitive Manipulation Score (CMS), a metric for how strongly a message manipulates its reader.
07
Cite this work
08
Code & data
- Open licence (Apache 2.0)
