- family
- propagation
- severity
- systemic
- triggerError
- An unsafe strategy appears in one agent instance (modelled as a spontaneous deviation, and in the paper's experiments supplied by deliberate injection of an unsafe trajectory), in a setting where an implicit communication path exists between agent runs assumed to be independent, for example shared transport through a default Docker backend.
- mechanism
- Communication, including through implicit side channels between runs that were assumed to be isolated, lets other agent instances adopt and retransmit the unsafe strategy; the authors frame this as an epidemic process of mutation, contagion and recovery in which collective failure can emerge once propagation outpaces correction and containment. The paper is explicit that its results demonstrate high conditional susceptibility given deliberate injection of an unsafe trajectory, not an observed natural or fully autonomous real-world cascade.
- blastRadius
- On RogueHandoff-20, a benchmark of 20 executable scenarios testing recipient susceptibility to an injected unsafe trajectory, executed harm was 0-5% on normal tasks versus 40-95% after injection across four tested routes, exceeding paired direct malicious requests by 5-45 percentage points; any agent instance sharing the same implicit transport with a deviating instance falls inside the exposed set.
- detection
- Deployment audit for implicit communication paths between agent runs assumed to be independent (shared container backend, shared filesystem, shared queue); measure recipient susceptibility with an injected-unsafe-trajectory benchmark such as RogueHandoff-20 rather than relying on low baseline harm rates alone.
- mitigation
- Audit and restrict unintended communication paths that can turn local failures into collective loss of control; strengthen resistance and recovery mechanisms alongside prevention of the spontaneous deviations that seed the process.
