Malicious prompt classifier

precision

98.84%

accuracy

90.79%

F1 score

89.96%

Published in Procedia Computer Science, 2026.

• Developed a probabilistic detector that scores a prompt by the transitions between its text sequences, modelled as a Markov chain.
• Added an explanation module that highlights the high-risk patterns behind each flag.
• Reached 90.79% accuracy, 98.84% precision, 82.54% recall and an 89.96% F1 score on a malicious-prompt benchmark.
• Published the paper in Procedia Computer Science (2026).

Chatbots can be tricked by prompts written to make them misbehave. This research catches those prompts before they get through, and says why each one was flagged.