Search the Atlas

Search risks, controls, and glossary terms

HighAgenticIncomplete Agent EvaluationDemonstrated

Validation Coverage Gaps

Accountability & Governance

Description

Standard model evaluations fail to capture the full behavioural space of agentic systems. Point-in-time testing insufficient for long-horizon agentic behaviour; multi-agent interaction risks not evaluated.

Example scenario

Agent passes all pre-deployment safety evaluations but exhibits harmful behaviour in production in unevaluated multi-step scenarios.

Real-world evidenceDemonstrated

Red-team evaluations of agentic LLM systems (including Anthropic's own published work on Claude tool use and DeepMind's evaluation of tool-using agents) have demonstrated that agents can invoke tools outside their intended scope when manipulated. No confirmed production incident of malicious tool abuse causing measurable harm in a regulated agentic deployment has been publicly attributed.

Primary mitigations

  • Continuous behavioural monitoring beyond static evals
  • multi-agent scenario testing
  • long-horizon evaluation frameworks
  • adversarial evaluation design.

Detection signals

Evaluation Coverage Score; behavioural space coverage; post-deployment drift from eval performance.

Mitigating controls

7
Dual coverage

Related risks in Accountability & Governance