AI Agent Security Assessment Method & Test Limits

1. Record the system version, tools, permissions, deployment and requirements.
2. Identify high-risk actions and an authorised test environment.
3. Agree scenarios, allowed inputs, attempt budgets, stop conditions and success criteria before testing.
4. Record each observed result and its supporting evidence.
5. Map findings to requirements and retest agreed fixes.

The engagement test plan carries its own version. This page does not invent a published benchmark version, default test budget or validated threshold.

Report successful and unsuccessful attacks as counts alongside attempted scenarios, covered actions and exclusions. Adaptive attempts may be related; a simple percentage is a descriptive summary of those tests, not a probability that the agent will fail in production.

No confidence interval is advertised until its statistical assumptions, unit of analysis and calculation have been established for the agreed test design. Findings use implemented, missing, failed in testing or not tested, rather than a numerical promise of certification.

This is invented example data, not a client test or a population risk estimate.

Repeat affected scenarios when models, prompts, tools, permissions or policies change, using a cadence and trigger process agreed with the client. Engineering checks performed by EndigitalX are first-party test results; independent assurance is a separate role. See Independence.

Testing covers the agreed actions and scenarios; it cannot enumerate every attack or replace privacy, safety and business-logic reviews. A favourable result does not guarantee future security, legal compliance or an external audit outcome.

Reviewers can request the applicable test plan and evidence under agreed confidentiality terms. No self-service testing portal or public attack-suite download is claimed.