4Claw AI Agent Dataset: 4,209 Real-World Prompt Injections

4,209 real-world prompt injection records were collected from the 4claw.org AI Agent Imageboard by David Keane at the National College of Ireland in March 2026. This cybersecurity-focused collection documents adversarial interactions and injection attempts within a multi-agent social media environment.

Use Cases

• Training text classification models to detect malicious prompt injections within agent-to-agent communications
• Analyzing the success of 'CyberRanger' security iterations by evaluating the 4,209 injection attempts
• Testing text generation safety by using the documented injection strings as adversarial inputs for LLMs

Strengths

• Contains 4,209 real-world prompt injection attempts
• Academic provenance from the National College of Ireland MSc Cybersecurity program
• Licensed under CC BY 4.0 for open research

Limitations

• Geographic bias toward the US region as indicated by metadata tags
• Domain-specific focus on imageboard-style interactions may not generalize to enterprise environments
• Potential for offensive or toxic language inherent to 4chan-style platforms

Provenance

Source

National College of Ireland and 4claw.org

Collection Method

Logged from the 4claw.org AI Agent Imageboard platform

Time Range

2026-03-06

Freshness

Collected in March 2026; reflects specific agent security versions available at that time.

Geography

United States