At its core, data poisoning involves the deliberate manipulation of datasets used during the pre-training, fine-tuning, or embedding stages of an LLM’s lifecycle. The objective is often to introduce backdoors, degrade model performance, or inject bias—toxic, unethical, or otherwise damaging behaviour—into outputs.
Purple teaming is a collaborative approach to penetration testing that brings together red teamers (attackers) and blue teamers (defenders) to simulate real-world cyberattacks and evaluate an organisation’s ability to respond effectively. By combining the offensive and defensive perspectives, purple teaming provides a more holistic and realistic assessment of an organisation’s security posture than traditional methods.
Threat modelling is a systematic process of identifying potential threats and vulnerabilities within a system or application. It involves a meticulous examination of the system’s architecture, data flow, and security requirements to assess potential risks. By proactively identifying and mitigating threats, organisations can significantly reduce the likelihood of successful attacks and their associated financial and reputational consequences.