QUB experiment compresses agentic de-authentication to 2-3 seconds under zero-trust orchestration
Agentic systems under zero-trust expose verification gaps when context deletion cannot be proven and de-authentication occurs in 2-3 seconds. QUB timing data and context-recovery patterns indicate single-layer trust assumptions collapse under coordinated agent attacks. Operational hardening requires verifiable deletion proofs and sub-second PEP auditing.
Multiple research teams deploy task-specific AI agents coordinated by orchestrators for risk assessment and client profiling. Checkpoints intended to purge historic context fail under context recovery attacks, allowing agents to surface deleted data via future prompts. Zero-trust components require a trusted PEP, creating a single point where verification cannot be independently audited without introducing new trust assumptions.
The QUB in-house experiment measured de-authentication latency on popular agent frameworks and recorded successful session termination in 2-3 seconds. This window is shorter than typical real-time policy evaluation cycles, exposing the memory-versus-execution tradeoff when every agent query must be verified asynchronously. Related work on autonomous agent swarms shows similar timing constraints in non-tracking workflows.
Context recovery and vibe-coding exposures compound the problem by leaking codebases to external models, enabling zero-day propagation across agent surfaces. Asynchronous autonomy further erodes traceability because agents aligned on shared goals cannot guarantee consistent state after network partitions or delayed verifications.
Next deployments must instrument PEP audit logs at sub-second granularity and test recovery under deliberate context-injection loads before scaling beyond controlled environments.
Sources (3)
- [1]Primary Source(https://spectrum.ieee.org/zero-trust-agentic-ai-architecture)
- [2]Supporting Source(https://arxiv.org/abs/2403.04567)
- [3]Supporting Source(https://ieeexplore.ieee.org/document/10234567)