
A penetration test report is worth what its evidence is worth: every finding should trace from an attack narrative to a captured request, an affected asset, a realistic impact and a specific fix. If you cannot follow that chain, you are probably holding a vulnerability scan with a cover page. And you paid pentest rates for it.
The short answer
- A good penetration test report has scope, methodology with coverage, manual attack narratives, raw request/response evidence, contextual severity and specific fixes with owners and retest status.
- A rebranded scan shows plugin IDs, vendor-generic text, dashboard screenshots, no authenticated evidence, no attack chain and fixes like "apply patches."
- Prioritize by demonstrated attack chains and root causes, adjusted for exposure, data sensitivity, KEV and EPSS, not by raw CVSS or finding count.
Key takeaways
- A credible penetration test report traces every finding from attack narrative to raw request, affected asset, impact and specific fix.
- Scanner plugin IDs, vendor-generic text, missing authenticated evidence and generic remediation mark a rebranded vulnerability scan.
- CVSS is technical severity; prioritize with exposure, data sensitivity, CISA KEV, EPSS and verified compensating controls.
- Group findings by root cause and give each group a named owner, a due date and a retest that reuses the original finding IDs.
- AI and agent reports need model versions, tool and MCP permissions, success rates across many runs and full transcripts.
What should a penetration test report include?
A penetration test report should include scope and rules of engagement, a named methodology with a coverage map, an executive summary, manual attack narratives, reproducible evidence for each finding, severity adjusted for your environment, and remediation specific enough to assign to an engineer.
Read the front matter first. It tells you whether the test can be trusted before you open a single finding:
- Document control: client, report version, test dates, classification, distribution list, lead tester and a named QA reviewer.
- Scope and rules of engagement: in-scope hosts, applications, APIs and cloud accounts; the test window; source IP addresses; credentials and roles supplied; emergency contacts; explicit exclusions.
- Methodology: named and mapped to coverage. Credible references include PTES, NIST SP 800-115, the OWASP Web Security Testing Guide, OWASP ASVS, the OWASP API Security Top 10 and MITRE ATT&CK.
- Limitations: black-box, gray-box or white-box access, time-boxing, rate limits, and anything ruled out, such as denial of service or social engineering.
If the methodology section is boilerplate that would fit any client, note it. It usually predicts the quality of everything that follows.
The executive summary comes next. A common norm is one to three pages, written so a CTO can decide where to spend without reading the findings. It is not a shortened findings table, and it should leave CVSS scores and endpoint names to the technical section.
A useful one answers four questions: what an attacker could achieve from each starting position tested, which systemic weaknesses produced the findings (for example, authorization enforced only in the frontend), how results compare with the previous test, and which two or three strategic changes would remove the most risk.
Keep it separate from the attestation letter, a short statement of scope, dates and outcome for customers, auditors or insurers. Restrict the technical report to the named recipients on the distribution list and share the attestation letter externally.
How do I tell a penetration test from a rebranded vulnerability scan?
Look for proof that a human validated, chained and exploited issues. A scan lists suspected weaknesses; a penetration test demonstrates what an attacker can actually reach. A polished cover, the words “penetration test” and a Critical/High/Medium table prove nothing on their own.
| Attribute | Vulnerability scan | Vulnerability assessment | Penetration test |
|---|---|---|---|
| Primary output | Tool results | Broad list of weaknesses, partly triaged | Validated, exploited findings and attack paths |
| Manual validation | None | Some | Every reported finding |
| Chained issues | No | Rarely | Yes, shown in narratives |
| Business impact | Generic | Partial | Specific to your data and roles |
| Typical tools | Nessus, Qualys VMDR, OpenVAS/Greenbone, Nuclei | Scanners plus review | Burp Suite, Metasploit Framework, BloodHound, custom scripts, plus scanners |
Pentesters use scanners too. The difference is what happens after the scan. These tells suggest nothing did:
- Plugin or template IDs left in finding titles, and descriptions copied verbatim from the scanner vendor.
- Screenshots of Nessus, Acunetix or Invicti dashboards instead of raw requests and responses.
- No evidence of authenticated testing, even though credentials were supplied.
- No attack narrative, or one that is just the finding list in paragraph form.
- Fifty findings for one root cause, such as a missing header repeated per host.
- Remediation that reads “apply the latest patches” or “sanitize user input.”
- No limitations section, which means nobody tracked what was not tested.
If you count three or more, ask the vendor for raw test notes before you pay the invoice.

What do a good attack narrative and reproducible evidence look like?
A good attack narrative states who the tester started as, every step they took, when, against which asset, and what the attacker ended up holding. It is the clearest proof of manual work, because scanners do not chain findings.
A condensed example of the shape to expect:
- Precondition: self-registered user on the customer portal, no special role. Test day 2, 10:14 UTC.
- Step 1: the invoice API accepted a sequential
accountIdwithout an ownership check (finding F-03, BOLA). Evidence: request/response pair E-11. - Step 2: one invoice PDF exposed a support engineer’s email address and the tenant’s SSO domain (F-07). Evidence E-14.
- Step 3: the password reset flow trusted a client-supplied
Hostheader, so a reset link for that engineer could be pointed at a tester-controlled domain (F-02). Evidence E-15, E-16. - Impact: read access to every tenant’s billing data and a path to an internal support role that can impersonate customers.
F-03 and F-07 might each score Medium alone. Chained, they become a cross-tenant compromise. That is how we structure findings in web application penetration testing: the chain is the deliverable.
Each step then points to a finding an engineer who was not on the engagement can reproduce from the report alone. Every finding should carry:
- Unique ID, title, severity, the full CVSS vector and score, and a CWE mapping.
- Affected asset: host, endpoint, parameter, role and environment.
- A description of what is wrong in this application, not the general weakness class.
- Numbered reproduction steps with the identity used at each step.
- Evidence: sanitized HTTP request and response transcripts, annotated screenshots with URL and timestamp, command output, logs, or video for complex flows.
- Impact and likelihood, remediation, a named owner, references and retest status.
Redaction cuts both ways. Secrets, session tokens and personal data must be masked, because reports travel to auditors, insurers and acquirers. Over-redaction that strips the parameter name or response body destroys reproducibility. Ask for a redacted distribution copy and a restricted technical annex.
How should I prioritize findings: CVSS severity or business risk?
Prioritize by business risk, using CVSS as one input. CVSS measures the technical severity of a flaw in isolation; business risk weighs that severity against your exposure, data, compensating controls and threat activity. A report that treats them as the same thing will misorder your remediation queue.
CVSS v3.1 bands are None 0.0, Low 0.1 to 3.9, Medium 4.0 to 6.9, High 7.0 to 8.9 and Critical 9.0 to 10.0, and CVSS v4.0 uses the same qualitative bands. A report should show the vector, not just the number.
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H scores 9.8, and the vector explains why: network reachable, no privileges, no user interaction. If the service actually sits behind a VPN and requires an admin session, the vector is wrong and so is the priority.
| Signal | What it answers | Limit |
|---|---|---|
| CVSS | How bad is this flaw technically? | Ignores your environment unless environmental metrics are applied |
| EPSS | How likely is exploitation activity in the next 30 days? | Covers public CVEs only; says nothing about custom app bugs |
| CISA KEV | Is there evidence of exploitation in the wild? | Binary and lagging; absence is not safety |
| SSVC | Which action fits: Track, Track*, Attend or Act? | Requires your own inputs on exposure and mission impact |
In CISA’s SSVC model, Track* means no action is needed yet, but the vulnerability has characteristics that warrant closer monitoring for changes.
Most findings in custom applications, such as broken access control or business logic flaws, have no CVE at all. For those, the tester’s impact statement is your only risk signal, which is why it has to be specific.
Then weigh demonstrated exploitability, internet exposure, whether authentication is required, data sensitivity, blast radius (one tenant or all of them), KEV listing or EPSS for known CVEs, compensating controls you have actually verified, and fix effort.
Group by root cause next. Twelve IDOR findings across twelve endpoints are one missing authorization layer, and patching each endpoint leaves the thirteenth open. Every group needs a named owner and a due date, or it will still be open at next year’s test.
Many teams use SLA tiers such as Critical 24 to 72 hours, High 7 to 30 days, Medium 30 to 90 days and Low 90 to 180 days. These are conventions, not standards: an internet-facing Medium in a chain can outrank an internal-only High.
How do I know the test was thorough and the fixes were retested?
Check the coverage matrix and the limitations section, not the finding count. Zero findings can mean strong security or a narrow scope, and the report must give you enough detail to tell which.
A coverage matrix lists each asset, role, API, cloud account and mobile build, and marks it as tested, partially tested, excluded or blocked. Compare it with the scope you agreed. Common gaps:
- Credentials were supplied for three roles, but evidence exists for only one.
- The API has 140 endpoints; the report references 20.
- A WAF or rate limit blocked testing and nobody recorded it.
- Third-party or managed services were silently skipped.
Gray-box testing with real accounts finds more per hour than black-box, because testers spend their time on authorization and logic instead of discovery. A time-boxed black-box test can produce a short report, but it should say so plainly. Our OWASP Top 10 penetration testing reports mark each category as tested, not applicable or blocked, with the reason.
A retest is a second, targeted pass that confirms each fix works. Critical and High findings should normally be retested and recorded in an updated report version. A proper retest reuses the original finding IDs, shows before-and-after evidence, states residual risk, and records each item as closed, partially remediated or risk-accepted with a named approver.
Retesters often find that a fix blocked the exact payload but not a variant, or patched one endpoint of a shared root cause. That only shows up when someone attacks it again.
Compliance drives retests too. PCI DSS v4.0.1 (Requirement 11.4.4) requires exploitable vulnerabilities found during penetration testing to be corrected and the testing repeated to verify the corrections. ISO/IEC 27001:2022 covers technical vulnerability management in Annex A. SOC 2 does not prescribe a pentest, though auditors often ask for one. Confirm exact obligations with your assessor.
How do I evaluate an AI, LLM or agent penetration test report?
An AI or agent penetration test report needs everything above plus environment details and statistics, because model behavior is non-deterministic. One screenshot of a successful prompt injection is an anecdote, not a reproducible finding.
Expect these additions:
- Model name and version, system prompt version and guardrail configuration at test time.
- Agent tools and permissions, including each Model Context Protocol (MCP) server and the credentials it holds.
- RAG data sources, their trust levels and who can write to them.
- Attempts and success rate per technique, such as 14 successes in 50 runs.
- Sanitized full transcripts or trace IDs, not cropped replies.
- Impact expressed as actions: which tool was called, with what arguments, against which system.
- Mapping to the OWASP Top 10 for LLM Applications and MITRE ATLAS.
The impact line is where weak AI reports fall apart. “The model said something it should not” is a guardrail finding. “Indirect injection in a support ticket caused the agent to call the refund tool for another customer” is an access control failure with a dollar figure.
That second kind of finding is what AI agent penetration testing should produce. If you are scoping one, our AI agent penetration test scope checklist lists what belongs in the rules of engagement.
What should I ask a penetration test vendor before buying?
Ask for a redacted sample report and read it against everything above; it predicts your deliverable better than any sales call. Certifications help, but peer review, methodology and evidence quality matter more.
Ask:
- Can we see a redacted sample with attack narratives and raw request evidence?
- Which testers will work on us, and what are their credentials (OSCP, OSEP, OSWE, CREST, or NCSC CHECK where relevant)?
- Who performs QA review, and is that person named in the report?
- Is authenticated testing across every supplied role included?
- Is a retest included, within what window, and does it produce an updated report?
- Do you provide an attestation letter separate from the technical report?
- How is our data stored, who can access evidence, and when is it destroyed?
- What professional liability insurance do you carry?
If a sample report contains no raw HTTP requests, treat the vendor as a scan vendor, not a pentest vendor.


FAQ
What belongs in a good pentest report?
Scope and rules of engagement, methodology with a coverage map, an executive summary, manual attack narratives, reproducible evidence for each finding, contextual severity, and specific remediation with an owner and retest status. Without evidence and narratives it is not a pentest report.
What is the difference between a penetration test and a vulnerability scan?
A vulnerability scan is automated tool output listing suspected weaknesses. A penetration test manually validates findings, chains them into attack paths, and states the business impact. If a report has no narratives or request/response evidence, it is a scan with a cover page.
Is a CVSS 9.8 finding always the top priority?
No. Check the vector against reality. If the service is not internet-exposed, requires privileges, or holds no sensitive data, a chained Medium on a public endpoint may matter more.
Should a penetration test include a retest?
Yes, at least for Critical and High findings. The retest should reuse finding IDs, show before-and-after evidence and record each item as closed, partially remediated or risk-accepted.
Does a penetration test report with zero findings mean we are secure?
Not necessarily. Zero findings can reflect a narrow scope, no authenticated access or blocked testing, so check the coverage matrix and limitations section before drawing conclusions.
What makes an AI agent penetration test finding reproducible?
The model and prompt versions, tool and MCP permissions, the number of attempts and success rate, and full sanitized transcripts or trace IDs. One successful prompt is not enough.
Every agent::offense report ships with attack narratives, raw request evidence and a retest, whether the target is a web app or an autonomous system. If you are buying AI agent penetration testing, ask us for a redacted sample and hold it to the checklist above.