Skip to content
Back to blog

Penetration Testing Rules of Engagement: Scope, Authorization and Stop Conditions

7 min read
Penetration Testing Rules of Engagement: Scope, Authorization and Stop Conditions

Sources checked: . FailSafe Research.

Rules of engagement are the part of a penetration test that tells an engineer what they may do at 2am without calling anyone. A signed proposal and a target list are not enough on their own. The record has to name who authorized the work, which systems and identities are covered, what effects are permitted, what must stop the run, and who decides whether work resumes.

This guide describes what a usable agreement contains. It is a planning aid, not legal advice. Confirm authorization, notice obligations and any third-party permissions with the responsible owner and your counsel.

Start from the authorization, not the target list

The authorization should come from someone who can grant it for the entity that owns the system. Record the legal entity, the signer and their role, the named testers, the testing window, and the systems covered. If the target depends on a hosting provider, a payment processor or another vendor, confirm whether that provider's policy requires separate permission before testing. A customer cannot authorize testing of infrastructure they do not control.

Keep the authorization with the test plan. If the scope changes during the engagement, update both. A verbal approval to add a hostname is difficult to defend later, and it is difficult for the tester to prove.

Define the boundary in terms a tester can check

An engineer should be able to look at a request and decide whether it is in scope. That means the boundary is written as concrete values, not categories:

  • Hostnames, IP ranges, API base paths, mobile app builds and cloud accounts that may be tested.
  • Environments covered, such as staging, pre-production or a named production slice.
  • Test identities and roles, including which tenant, workspace or customer record each one represents.
  • Explicit exclusions, including third-party services, data stores and endpoints that must not be touched.
  • Shared infrastructure that requires coordination, such as load balancers, queues or monitoring that other teams depend on.

The target list should be specific enough that a request to an unlisted host is visibly out of scope. "The application" is not a boundary.

Record the permitted effects

Testing methods produce different effects. Agree which ones are allowed rather than assuming a generic label covers everything.

Rules-of-engagement elements and the check that confirms them
ElementWhat to recordHow to confirm it
AuthorizationSigner, entity, testers, dates and covered systems.A written approval that matches the final scope.
TargetsHosts, paths, environments, accounts and exclusions.A test request to an unlisted host is rejected in staging.
IdentitiesTest accounts, roles, tenants and synthetic records.Each account can be disabled at close-out.
MethodsDiscovery, validation, exploitation, persistence and social engineering, each marked allowed or excluded.An excluded method is not present in the run configuration.
Safe effectsRead-only defaults, request limits, destructive action bans and test-data rules.Limits appear in the tooling configuration, not only the document.
Data handlingWhat may be captured, where evidence is stored, who can read it and when it is deleted.Evidence location and retention date are named.
CommunicationsPrimary and escalation contacts on both sides, channel and response window.A test message reaches the named contact.
Stop conditionsEvents that pause the run and the person who decides to resume.The stop path is exercised before the first live action.
Close-outAccount removal, token revocation, artifact cleanup and confirmation.A signed close-out record with the actions performed.

Separate discovery from exploitation

Many engagements allow broad discovery and narrow exploitation. Write that distinction down. For example, enumeration of an API may be permitted across the agreed host, while proving a cross-tenant read is limited to two synthetic records and one named test identity. Persistence, lateral movement, denial of service and social engineering are separate decisions and should not be implied by the word "pentest".

Where a consequence is difficult to demonstrate safely, agree a bounded proof: a read of a single synthetic record, a request that returns a permission decision, or a log entry showing the attempted action was denied. The proof should show the boundary without exposing real customer data or degrading the service.

Write stop conditions as instructions

A stop condition is only useful if the tester knows what to do when it occurs. Each one should name the trigger, the immediate action, the contact and the person who decides whether work resumes.

Common stop conditions and the response they should trigger
TriggerImmediate actionResume decision
Real customer data appears where synthetic data was expectedStop the action, preserve the observed state without copying the data, notify the owner.Data owner and engagement lead after the cause is understood.
Service degradation or an unexpected outageStop testing against the affected component and notify the on-call contact.Service owner confirms stability.
An unapproved side effect occursHalt the run, record the action and its result, do not repeat it.Written approval before any further action on that path.
A test identity reaches data outside the agreed boundaryStop, record the access path, revoke the identity if needed.Scope review before testing continues.
Credentials or secrets are exposed in evidenceStop capture, rotate through the owner's process, store only a redacted reference.Owner confirms rotation.
The target differs from the agreed configurationPause and confirm which build is deployed.Scope and configuration re-agreed.

Give the tester authority to stop without waiting for approval. Resuming is the decision that needs a named owner.

Agree how evidence and communications work

Name the channel for findings, the person who receives them, and the expected response time for a critical issue. A finding that sits in a shared folder over a weekend is not an escalation path. Agree how a suspected incident is handled: who calls whom, what the tester may say to other teams, and whether the engagement pauses.

Evidence should be sufficient to reproduce the finding and no more. Store raw evidence in the agreed location, share sanitized extracts with wider reviewers, and record where the original is retained. If logs, traces or telemetry are unavailable, label the conclusion as limited rather than filling the gap with inference.

A fictional walkthrough

A staging invoice application has a browser interface, an API and three roles. The owner authorizes two test identities per role, a staging host and a read-only API token. The agreement permits enumeration, validation and a bounded cross-tenant read of one synthetic invoice. It excludes persistence, denial of service and any use of real customer records.

During the test, the second Reader account returns a record that appears to be a real invoice. The stop condition applies. The tester halts, records the request and the response metadata without copying the record, and notifies the owner. The owner confirms that a restored production snapshot was loaded into staging by mistake. Staging is reset, the synthetic record is replaced, and the engagement lead approves a restart on the corrected build.

The finding is recorded with the environment mismatch, the action, the observed boundary and the restart approval. The retest repeats the same case on the corrected data set and keeps the permitted same-tenant read as a control. The record shows what happened and who decided to continue, which is the point of writing the rules down.

Close-out is part of the engagement

At the end, disable test accounts, revoke tokens and keys, remove test records and any tooling left on the target, and confirm with the owner that the cleanup is complete. Record the close-out alongside the findings. If remediation retesting is included, state which original cases will be repeated, the window, and what counts as a pass. A merged code change is not the same as a confirmed retest.

Where FailSafe fits

FailSafe scopes authorized testing with written authorization, target limits, permitted actions, stop rules and an evidence plan before execution. The rules-of-engagement checklist records the elements above. The quote comparison worksheet keeps scope and effort visible when comparing proposals, and the synthetic finding and retest record shows the evidence trail a bounded retest should produce.

For agentic testing, the agentic pentesting guide describes the controls that apply when autonomous agents perform the work. MCP and AI assessments use the MCP scope worksheet and the AI assessment workbook. We do not test systems we are not authorized to test, and we do not publish working payloads or reproduction steps.

Questions & answers

Questions about this topic

Direct answers to common questions about this page.

The authorization and signer, target boundary, test identities, permitted and excluded methods, safe effects, data handling, evidence location, contacts, stop conditions and close-out steps.

Someone who can grant permission for the entity that owns the system. Name the signer, their role and the covered systems in writing. Confirm any separate permission required from hosting or service providers.

Real customer data appearing in a synthetic test, service degradation, an unapproved side effect, access outside the agreed boundary, exposed credentials, or a target that differs from the agreed configuration.

No. A contract sets commercial terms. Rules of engagement translate the agreed work into operational limits an engineer can follow during the test.

Only with explicit authorization and safeguards agreed for that environment. Many teams start in staging and permit a narrow production check with request limits, synthetic data and a named on-call contact.

Put this guidance to work

Bring the targets, identities and stop rules you need. FailSafe can review the boundary and agree the evidence before testing.

Startup offer: $999 USD for 250 credits and one authorized black-box run. Sign in to SWARM before Stripe checkout. Custom scope and sovereign deployments are separate.