Opaque Reasoning Chains
Accountability & GovernanceDescription
Stakeholders cannot inspect agent reasoning, tool invocations, or memory state. Explainability gap in autonomous multi-step decision making prevents effective oversight, debugging, and accountability.
Agent makes a consequential hiring recommendation with no explainable rationale, preventing HR from reviewing its basis.
While supply chain compromise is a Realized risk in software broadly (SolarWinds, XZ Utils), a confirmed incident in which an AI agent framework, plugin, or third-party model component was deliberately compromised and exploited in a production agentic deployment has not been publicly documented. The attack vector is plausible by analogy but undemonstrated for AI-specific agent supply chains.
No public incident on record — evidence level: Theoretical
Primary mitigations
- Reasoning trace logging
- step-by-step decision explanation
- human-readable audit trails
- interpretability tools integrated into agent pipeline.
Detection signals
Transparency Coverage Score; reasoning trace completeness rate.