Clawk AI Agent: 4,209 Prompt Injections for Cybersecurity Hardening

4,209 real prompt injection records collected by David Keane at the National College of Ireland in March 2026. The data documents the security evolution of AI agents from basic chatbots to the CyberRanger V42 Gold model.

Use Cases

• Training text classification models to identify malicious prompt injections
• Benchmarking multi-agent system security against adversarial text inputs
• Analyzing the success rates of 4,209 injection attempts across 42 agent versions

Strengths

• Contains 4,209 real-world prompt injection attempts
• Documents 42 distinct versions of agent security evolution
• MIT licensed for open cybersecurity research

Limitations

• Specific to the Clawk and CyberRanger agent architectures
• Limited to text-based prompt injection attacks
• Geographic scope is likely restricted to the research environment

Provenance

Source

National College of Ireland

Collection Method

Experimental research involving 4,209 injection attempts

Freshness

Last updated March 2026

Geography

Ireland