Researchers at Queen’s University Belfast are examining the security implications and the viability of zero-trust architectures in autonomous agentic AI systems. While teams globally are increasingly using AI agents for tasks like risk assessment and decision-making, experts note that these agents might fail to follow data deletion instructions, potentially exposing sensitive information through context recovery attacks.
An in-house experiment demonstrated that agentic orchestration could drastically reduce the time frame of cyber threats. The research team found that popular de-authentication attacks could be executed in as little as 2 to 3 seconds when driven by autonomous agents. This rapid pace could leave system administrators with almost no time to react, patch vulnerabilities, or deploy defensive measures.
Additionally, the researchers warn that practices like vibe coding—using AI models to assist in software development—can expose entire codebases to untrusted external entities, leading to widespread vulnerabilities. Experts point out that a single rogue agent might cause sweeping failures across a supply chain, and recovering from such asynchronous autonomous system breakdowns could be computationally expensive and difficult to contain.
To address these traceability and trust issues, the research group suggests utilizing concepts like an AI-bill of materials or distributed ledger technology. However, a fundamental question remains open for the researchers: whether a true zero-trust architecture can even exist in highly autonomous, stateless AI systems where trust-provenance decreases significantly.