Search the Atlas

Search risks, controls, and glossary terms

MediumAgenticOversight Saturation AttacksDemonstrated

Governance Attention Exploitation

Accountability & Governance

Description

Adversary generates excessive alerts, false positives, or trivial events to exhaust human oversight capacity, creating blind spots for real attacks. Exploits limited human attention in multi-agent monitoring environments.

Example scenario

Adversary triggers thousands of minor violations to overwhelm the security team while a major data exfiltration occurs undetected.

Real-world evidenceDemonstrated

Academic studies have shown that GitHub Copilot and other code-generation LLMs produce insecure code at statistically significant rates (NYU/Stanford research found ~40% of Copilot suggestions in security-relevant contexts contained vulnerabilities). Confirmed production incidents where agent-generated code was executed and caused a security breach have not been publicly attributed specifically to AI-generated code.

Primary mitigations

  • Alert deduplication and prioritisation
  • automated triage of low-risk events
  • attention economics monitoring
  • human reviewer workload management.

Detection signals

Alert noise ratio; false positive rate; oversight saturation metrics.

Mitigating controls

7
Dual coverage

Related risks in Accountability & Governance