Agentic AI built on large language models can perform multi‑step planning, invoke external tools, retain information in memory and coordinate with other agents. These capabilities increase usefulness but also introduce security and operational risks such as untrusted content from websites, emails, documents or databases mixing with system instructions, persistent memory carrying compromised data across sessions, and tool access turning incorrect responses into real‑world actions.
Trustworthiness is examined across five interrelated dimensions: safety and robustness, alignment with human oversight, transparency and auditability, privacy and data governance, and regulatory compliance.
Key failure modes are organized into a taxonomy that includes indirect prompt injection, backdoor triggers, goal misgeneralization, memory contamination, and cross‑session data leakage.
Major mitigation strategies comprise instruction hierarchies, context isolation, spotlighting, process‑based supervision, constrained tool use, and privacy‑preserving memory. The article distinguishes techniques with empirical support from those that remain largely conceptual.
Building on this analysis, the Trustworthy Agent Development Lifecycle (TADL) is introduced, covering specification, design, training, evaluation, deployment and monitoring. For each phase it identifies relevant trust activities, expected evidence and risk‑based decision gates.
Although TADL has not yet been empirically validated, it offers a structured foundation for developing more secure and accountable agentic systems. The paper concludes by highlighting gaps in current benchmarks and outlining priorities for future research.
Review