AI Agents Are Testing the Limits of Zero-Trust Security

Vishal Sharma, a Reader in the School of Electronics, Electrical Engineering and Computer Science at Queen's University Belfast, sets out in an IEEE Spectrum piece a set of problems that emerge when autonomous agents are chained together under zero-trust architectures. The core observation is a circularity: zero-trust means "never trust, always verify," yet the verification runs through a policy enforcement point (PEP) that must itself be trusted for the rest of the chain to hold. Sharma asks directly where the zero-trust element resides if the policies enforced by the PEP are themselves taken on faith.
One failure mode he describes involves memory rather than access control. Teams building orchestrator-led agent fleets for risk assessment, client profiling, match making and information gathering prompt agents to periodically delete historical context so the system does not skew toward repeating earlier decisions. A later prompt can nevertheless cause an agent to restore material it was supposed to have discarded, a context recovery attack. That undercuts the assumption that deletion equals forgetting, and raises the broader accountability question of how an operator proves an instruction was followed at all.
A concrete measurement comes from an in-house experiment at QUB. Focusing on de-authentication attacks, the sort used to knock a device off its network, the team found that with agentic orchestration the pace of such attacks could drop to as low as 2 to 3 seconds. The described scenario involves an agent scanning an active channel and autonomously intercepting traffic against an aerial vehicle, compressing the window in which a defender can react or patch to nearly nothing. In practice this reframes de-authentication from a nuisance flooding technique into a fast, automated denial-of-service primitive.
Sharma also flags supply chain and tooling risk. He points to vibe coding, where a model is handed source code to help write software, as an example of exposing a full codebase to an untrusted external entity, with vulnerabilities potentially spreading across multiple surfaces. He argues the scale question cuts both ways: in a dense web of agents operating with low tolerance for failure, verifying every query in real time is computationally expensive, forcing a trade-off between memory and execution.
Resilience is another theme. With only one trusted layer, a zero-day or a loss of the backbone network removes the place where trust is anchored, and the whole ecosystem degrades at once. Under asynchronous autonomy, where agents pursue shared goals on different clocks, Sharma treats trust as the basis for traceability and asks how alignment survives those delays. His group's proposed mitigations lean on artifacts that make agent behavior auditable and platform-agnostic: an AI bill of materials, and distributed ledger technology for traceable, immutable records of agent activity and its trust vectors. He also names horizontal anomalies, where an agent that behaves correctly in one task or application cannot be assumed trustworthy across all the components it touches, as an unresolved detection problem.
On recovery, Sharma argues that low-trust agent deployments can deliver short-term benefit while the cost of a failure remains unmeasured and unlikely to stay contained. A rogue agent surfacing through a supply chain becomes a problem spanning every touch point, so even proxy agents only loosely coupled to the wider system need trackers, expensive as those are to run.
For organizations where trust is a prerequisite, Sharma describes a posture of internal policy guardrails rather than trust in third-party agentic providers, with activities tracked through logs and token revocation lists used against replay attacks. He pairs this with formal verification of model robustness, weighing adversarial accuracy against certified adversarial robustness to estimate how far a model can be trusted. Related PhD work in his group addresses AI sovereignty and accountability, distinguishing missed information from missed instructions.
This suggests the practical difficulty is that agent security inherits the classic problem of a trusted computing base: every verification layer adds latency and cost, and the root of trust is still a single component someone must vouch for. Stateless designs with high agent autonomy reduce the state a defender can rely on, which plausibly erodes provenance. The QUB de-authentication result is a reminder that agent speed advantages accrue to attackers as readily as defenders.
Why it matters: Teams deploying multi-agent systems for profiling, assessment or code generation are being asked to treat every agent interaction as untrusted while accepting that the enforcement layer, logs and deletion checkpoints are all attack surfaces. Expect scrutiny of policy enforcement points, memory handling and third-party agent providers, and budget for continuous audit rather than one-time review. Whether an AI bill of materials or ledger-backed logging becomes standard practice is an open question, but the trade-off between verification cost and exposure is now a design decision rather than a compliance checkbox.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
This article is brought to you by Queen’s University Belfast . Several teams around the globe are leveraging AI by creating a task force of agents managed by an orchestrator and creating applications to support risk assessment, client profiling, match making, and information yield with a slight overview of decision making. One of the important components for these agents is to ensure that the logs are managed and historic relevance is not being used to avoid skewing towards same decision making. Thus, the system is