Moltbook AI Injection Dataset: 4,209 Real Adversarial Prompts for Cybersecurity

4,209 real AI-to-AI prompt injections collected by researcher David Keane for an MSc Cybersecurity project at the National College of Ireland in February 2026. The dataset documents the adversarial testing journey from a 'dentist chatbot' to a defensive system called 'CyberRanger V42 Gold' with a 100% block rate.

Use Cases

• Training and evaluating AI defense models based on the documented real adversarial prompts.
• Studying the psychology and engineering of prompt injection attacks based on the description.
• Benchmarking the robustness of conversational AI systems against the 4,209 documented injections.

Strengths

• Contains 4,209 real-world adversarial prompts.
• Documents a specific defensive outcome: a system ('V42 Gold') achieving a 100% block rate.
• Created for a formal academic research project (MSc Cybersecurity).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Researcher David Keane (IR240474), National College of Ireland.

Collection Method

Collected as part of an MSc Cybersecurity project, likely involving adversarial testing of AI systems.

Time Range

Collected in February 2026.

Freshness

Last updated 2026-03-09 02:56:46; freshness should be verified.

Geography

Geography is unknown.