Skip to content
CANHA

Methodology

How an assessment runs

Six phases, from agreement on scope to verified remediation. Every step is documented, contained, and reproducible.
  1. 01

    Architecture & Scope

    We start by understanding the agent: what it is, what it can touch, who uses it, and how the system is designed to control its access. A written scope — systems, environments, data, and guardrails — is agreed before any testing begins.

    • Review of agent architecture and permission model
    • Inventory of tools, APIs, integrations, and data stores
    • Written scope and rules of engagement

    Phase 01

  2. 02

    Attack Surface Mapping

    We enumerate every capability an adversary could abuse: the tools the agent can call, the data it can read, the actions it can take, and the identities it holds.

    • Capability and permission enumeration
    • Identification of injection and abuse vectors
    • Mapping of trust boundaries and likely failure modes

    Phase 02

  3. 03

    Adversarial Testing

    This is where we attempt to break intent. We try to make the agent exceed its authorization — through indirect prompt injection, tool misuse, authorization bypasses, and chains of individually low-risk actions.

    • Prompt and indirect injection testing
    • Tool and API misuse attempts
    • Authorization and boundary testing

    Phase 03

  4. 04

    Exploitation & Evidence

    Every confirmed finding is proven, not claimed. We demonstrate impact on the agent in its environment, capture the exact steps to reproduce, and record the chain of events.

    • Impact confirmed against real agent behavior
    • Step-by-step reproduction captured
    • Logging and traceability reviewed as part of each finding

    Phase 04

  5. 05

    Reporting

    You receive the executive summary for your leadership, the technical report for your engineers, and an evidence package built for your customer's security review. Findings are rated by real-world impact, with remediation guidance.

    • Executive security summary
    • Technical report with severity and impact
    • Evidence package for enterprise review

    Phase 05

  6. 06

    Retest

    After you apply fixes, we rerun the same scope, verify each remediation against the original findings, and issue an updated report.

    • Fixes verified against original findings
    • Updated report for your customer's review
    • Closing summary of what changed between runs

    Phase 06

Guarantees

Contained by design

Adversarial testing has to be trusted before it's useful. These are the constraints we operate under for every engagement.

Testing on your authorization

Scope and rules of engagement are agreed in writing before testing starts. Nothing outside that scope is touched.

Real behavior, safe environment

We exercise your actual agent and its integrations where you authorize it, and isolated fixtures and synthetic data where you'd rather we didn't touch production.

Evidence, not guesswork

No finding ships without the steps to reproduce it. If we can't demonstrate something, we don't report it as a finding.

We also keep clear of the tactics that produce noise without value: we don't social-engineer your users, we don't test third-party systems beyond written authorization, and we don't run unplanned or unbounded actions against your production environment.

An assessment your customer's security team will take seriously.

Request an Assessment