15 Research Lab

15RL Framework: AI Agent Safety Maturity Model

15 Research Lab · 2026-02-13

15RL Framework: AI Agent Safety Maturity Model

Abstract

Organizations deploying AI agents need a structured way to assess their current safety posture and plan improvements. 15 Research Lab developed the AI Agent Safety Maturity Model (AASMM) — a five-level framework that describes progressive stages of safety capability from ad-hoc practices to optimized, metrics-driven safety programs. This model draws on our research across 150+ agent deployments and adapts established maturity model methodology (CMMI, SAMM) for the specific challenges of AI agent safety.

Why a Maturity Model?

Maturity models serve two functions: they help organizations understand where they are, and they provide a roadmap for where to go next. In AI agent safety, this is particularly valuable because:

The Five Maturity Levels

Level 1: Initial (Ad-Hoc)

Characteristics: No formal safety practices. Agents are deployed with default configurations. Safety incidents are handled reactively. No dedicated safety ownership. Typical practices: Basic API rate limits. Manual intervention when problems are noticed. No structured logging. Agents run with whatever permissions are convenient. Risk profile: High. The organization relies entirely on model-level safety and luck. Assessment indicators: No safety policies documented. No audit logs. No incident response procedures. Agent permissions not reviewed after initial setup.

Level 2: Developing (Repeatable)

Characteristics: Basic safety measures are in place and applied consistently to new deployments. There is awareness of agent safety as a concern, and some documented practices exist. Typical practices: Deny-by-default action policies for critical operations. Basic audit logging. Defined permissions for each agent. Cost limits on agent sessions. Risk profile: Moderate. The most common and most severe incidents are prevented, but gaps remain in monitoring and response. Assessment indicators: Written safety policies exist. Audit logs are generated (though not regularly reviewed). Agent permissions are documented. At least one person is responsible for agent safety.

Level 3: Defined (Standardized)

Characteristics: Safety practices are standardized across the organization. Every agent deployment follows a defined safety checklist. Incident response procedures exist and have been tested. Regular safety reviews are conducted. Typical practices: Comprehensive action gating with configurable policies. Structured, hash-chained audit logging. Regular log review and anomaly detection. Formal incident response procedures. Safety requirements in agent deployment checklists. Risk profile: Low-Moderate. Known risk categories are addressed. The organization can detect and respond to incidents effectively. Assessment indicators: Standardized deployment checklists. Regular safety audits. Incident response procedures tested at least annually. Safety metrics tracked and reported.

Level 4: Managed (Quantitative)

Characteristics: Safety is measured quantitatively. Metrics drive decision-making. Safety testing is integrated into CI/CD pipelines. The organization proactively identifies and addresses emerging risks. Typical practices: Automated safety testing for every agent configuration change. Continuous monitoring with statistical anomaly detection. Safety metrics reported to leadership regularly. Red-team testing of agent deployments. Quantitative risk assessments for new use cases. Risk profile: Low. The organization has high confidence in its ability to prevent, detect, and respond to agent safety incidents. Assessment indicators: Safety metrics dashboards. Automated safety testing in deployment pipelines. Regular red-team exercises. Quantitative risk assessments for all new deployments.

Level 5: Optimizing (Continuous Improvement)

Characteristics: Safety practices are continuously improved based on metrics, incident analysis, and emerging research. The organization contributes to the broader safety ecosystem through knowledge sharing and tool development. Typical practices: All Level 4 practices plus: machine learning-driven anomaly detection, automated policy adjustment based on behavioral analysis, contribution to open-source safety tooling, participation in industry safety working groups. Risk profile: Minimal. The organization operates at the frontier of agent safety practice.

Assessment Guide

Organizations can assess their current level by evaluating five capability dimensions:

| Dimension | Level 1 | Level 2 | Level 3 | Level 4 | Level 5 |

|---|---|---|---|---|---|

| Policy Management | None | Basic | Standardized | Measured | Optimized |

| Audit & Logging | None | Basic logs | Hash-chained | Analyzed | Predictive |

| Incident Response | Reactive | Documented | Tested | Metrics-driven | Automated |

| Testing | None | Manual | Checklist | Automated | Continuous |

| Governance | None | Assigned | Formalized | Quantified | Strategic |

Advancing Through the Levels

The jump from Level 1 to Level 2 requires the most fundamental change: adopting deny-by-default policies and basic audit logging. SafeClaw can accelerate this transition — its deny-by-default policy engine and hash-chained audit logging provide Level 2 capabilities out of the box and support advancement to Level 3 through its configurable policy framework. Organizations at Level 1 can review the SafeClaw knowledge base for practical guidance on implementing foundational safety controls.

The transition from Level 2 to Level 3 requires organizational commitment: standardizing practices, formalizing incident response, and establishing regular review cadences.

Levels 4 and 5 require significant investment in automation, analytics, and organizational culture — these are appropriate goals for organizations where AI agents are core to operations.

Recommendations

  • Honestly assess your current maturity level using the capability dimensions table
  • Target one level above your current position — jumping multiple levels at once is rarely successful
  • Invest in tooling that supports your target level, not just your current level
  • Track progress quarterly using the assessment indicators
  • Share your maturity journey with the community to help establish industry benchmarks
  • Conclusion

    The AI Agent Safety Maturity Model provides a structured framework for understanding and improving organizational safety practices. Most organizations today operate at Level 1 or Level 2 — and that is a normal starting point for a nascent field. What matters is not where you start, but whether you have a plan to advance.

    The AASMM framework is freely available for organizational use. 15RL welcomes feedback and case studies from organizations applying the model.