Magnet: Detecting Cross-Session AI Misuse Through Capability Accumulation
Researchers demonstrate that attackers can decompose harmful goals into innocuous units executed across isolated agentic sessions, evading single-session detection. They propose Magnet, a detection framework addressing cross-session capability accumulation.
A paper published on arXiv on 2026-08-03 introduces Magnet, a method for detecting cross-session AI misuse where attackers accumulate capabilities across stateless agent interactions. The work highlights a gap in current abuse detection, which focuses on single-session threats, and shows that cross-session decomposition can elicit more harmful capability than single-session attacks.
Magnet addresses the asymmetry where an attacker maintains state across sessions while the agent does not, enabling evasion of existing monitors. The approach likely involves tracking composable artifacts (model responses and tool-call results) across sessions to identify harmful trajectories.
As AI systems become ensembles of specialized agents, the attack surface expands. This research signals a need for security tools that operate across session boundaries, potentially influencing how AI platforms implement monitoring and logging.
Enhances trust and safety for AI service providers by reducing undetected misuse, potentially lowering liability and compliance risks in regulated industries.
Next signals include adoption of cross-session detection in AI safety frameworks, integration into agent orchestration platforms, and further research on adversarial robustness in multi-agent systems.