Low-and-Slow Policy Bypass
Accountability & GovernanceDescription
Agent or adversary operates below detection thresholds—distributing actions across time, systems, and agent instances to obscure governance violations. Responsibility trails deliberately obscured across multi-agent architectures.
Adversary exfiltrates data at 1 KB/hour across 100 agents over 6 months, staying below all threshold-based detection systems.
Denial-of-service via resource exhaustion is a Realized risk for web services generally, but targeted saturation of agentic AI infrastructure as a distinct attack vector has not been demonstrated in published research or confirmed production incidents. The risk is architecturally plausible given agent loop overhead but remains undemonstrated empirically.
No public incident on record — evidence level: Theoretical
Primary mitigations
- Longitudinal behavioural analysis
- distributed action correlation
- cross-agent session linking
- anomaly detection at population level.
Detection signals
Governance Evasion Detection Rate; distributed action sequence analysis.