Authorization-first assessment and the KernelHawk Baseline score

Executive summary

A security score is useful only if a reader can answer five questions: What was in scope? When was it assessed? Which evidence supports it? How was it calculated? What important uncertainty remains?

KernelHawk Baseline separates commercial qualification from security assessment. Public business records and historical observations may help identify an organization that could benefit from a conversation. They cannot support a claim that the organization is vulnerable, breached, noncompliant, or unsafe. The customer-facing 1-10 score exists only after written authorization and evidence collection inside the agreed boundary.

The current score uses five required dimensions:

1. External exposure.
2. Identity and access.
3. Data protection.
4. Patch and configuration.
5. Detection and response.

Each dimension has four criteria, so a complete assessment is 20 out of 20. For every criterion an assessor records the state, the anchor that justifies it, the reasoning, and the exact source. The software does all the arithmetic — an assessor cannot type a number into the score. And if anything is missing (a criterion, a reference, an unlinked critical finding), the scoring fails rather than producing a number.

Reported cybercrime losses rose fivefold from 2020 to 2025

FBI Internet Crime Complaint Center annual data show a large increase in complaint-reported losses from 2020 through 2025. The series rises from $4.2 billion in 2020 to $20.9 billion in 2025, about five times the starting value. Complaint counts rose overall but not every year. The data support a measured increase in reported losses; they do not by themselves prove exponential growth, measure every cybercrime, or identify one cause.

AI acts as a force multiplier within the broader trend. The FBI's 2025 report says widely available AI can produce convincing synthetic profiles and personalized conversations in mass quantities, and that investment scammers can quickly generate thousands of distinct-looking conversations. IC3 recorded 22,364 complaints with the tracking descriptor "AI Related" and $893 million in adjusted losses in 2025. That label is a descriptor attached after a crime type is selected, so the figures provide context rather than a separate census of AI-caused crime.

The mechanism is consistent with official qualitative assessments. The June 22, 2026 Five Eyes statement, The AI Shift in Cyber Risk: Why Leaders Must Act Now, says AI lowers barriers for malicious actors, accelerates the speed, scale, and sophistication of cyber threats, and shortens the interval between vulnerability discovery and exploitation. KernelHawk therefore treats lower skill, time, and cost for producing code, preparing exploit attempts, personalizing lures, and operating at scale as a reason to reassess exposure quickly. That is an evidence-based inference about amplification, not a claim that AI alone produced the IC3 trend or that every AI-assisted attempt succeeds.

1. The authorization boundary

NIST SP 800-115 describes planning, conducting, analyzing, and using technical security tests and assessments. KernelHawk requires an agreed plan before assessment work begins. An engagement cannot proceed without:

• A named legal entity and authorizing person.
• A content digest of the signed agreement.
• A non-empty structured scope identifying targets and exclusions.
• Explicit start and end times.
• Active status and rules of engagement.
• An engagement identity that matches the authorized customer.

The authorization gate requires a signed scope tied to the customer and test window. Email, forms, public assets, sales notes, and payment cannot substitute for it. Expired, revoked, empty, or mismatched records are rejected.

1.1 Rules of engagement

The rules cover what we may do, from where, and when — permitted techniques, source addresses and maintenance windows, for example. They also name the fragile systems, the stop conditions, and who to call if we find unexpected access.

1.2 Scope changes

New assets or techniques are not absorbed informally. A scope change updates the written authorization, dates, rules, and commercial terms before work crosses the original boundary.

2. Evidence quality

The assessment record distinguishes evidence classes:

• Provided context: architecture, inventory, policy, or configuration supplied by the customer.
• Source observation: a bounded collector, scanner, artifact parser, or approved external source.
• Analyst validation: reviewed behavior or configuration supporting affectedness, reachability, control state, or impact.
• Reproduction evidence: a controlled demonstration within scope, with exact request, response, artifact, or state transition.
• Remediation evidence: a verified before-and-after observation after corrective work.

Evidence should identify source, time, target; method, reviewer where appropriate, and a stable reference or digest. Sensitive values are redacted or stored in a separately controlled evidence environment; credentials do not belong in the exposure database or public report.

Score model version 3 records the evidence used for all 20 checks and calculates the result from those records. Before a report is published, the software checks that its evidence and calculation still match. Changed evidence requires a new assessment record.

2.1 Coverage is part of the result

Missing data does not produce a polished score. If a required dimension cannot be assessed to the agreed depth, the model refuses to publish an overall number, or the report narrows the scope and says so. Unknown coverage counts as unknown, never as a control that passed.

3. The five dimensions

3.1 External exposure

Measures asset visibility, ownership, internet reachability, service minimization; secure edge configuration, external dependencies, and verified remediation of reachable weaknesses. Evidence may include asset and DNS inventory, service configuration, approved observations, certificates, and attack paths.

3.2 Identity and access

Measures whether trust can be established, limited, revoked, and audited — across your identity inventory, your privileged accounts, and your service accounts, among others.

3.3 Data protection

Measures data classification, collection minimization, access boundaries, encryption, and key ownership, backup protection; retention, deletion, leakage controls, and recovery requirements. Controls are evaluated against the data and business process actually in scope.

3.4 Patch and configuration

Measures inventory-to-update linkage, vulnerability prioritization, configuration baselines, exception handling; supported versions, software, and device lifecycle, cloud posture, and verification. CISA KEV and EPSS can inform order but do not replace local affectedness and impact.

3.5 Detection and response

Measures useful telemetry, alert coverage, triage, containment authority, incident roles; communication, restoration, exercises, lessons learned, and proof that critical services can recover.

3.6 Closed observation states

There is no “not applicable” and no “unknown.” Picking a state and arguing for it is not enough — the observation has to cite the exact published anchor for that criterion. Managed and verified states also need a reference into the evidence set the score is locked to.

4. Score model version 3

Each dimension is the equally weighted mean of its four state values, multiplied by 100. Open priority findings then apply a ceiling to every dimension: 60 when any critical is open, otherwise 85 when any high is open, otherwise 100. Let D be the arithmetic mean of the five resulting dimension values.

Base score = 1 + (D x 9 / 100)

Final score held between 1.0 and 10.0, rounded half up to one decimal place.

• 1.0 point for each open critical finding.
• 0.25 point for each open high finding.
• Total penalty capped at 4.0 points.
• Final score clamped to 1.0-10.0 and rounded half up to one decimal place.

The critical and high counts are read back from the same evidence set the score is locked to, and the individual findings have to match those counts. Every one of them must be referenced by at least one scored criterion. Nothing from prospect research or old public signals can reach this score.

Why both a ceiling and a penalty?

The ceiling prevents a perfect-looking dimension profile from coexisting with an unresolved priority condition. The separate penalty makes the count and urgency visible in the final 1-10 result. The detailed register still carries every finding.

5. Worked example

Consider an illustrative 20/20 assessment whose four-state calculations produce raw dimension values of 90, 65, 77.5, 65, and 65. The evidence contains one open high finding. The 85 ceiling changes the first dimension from 90 to 85, producing a dimension mean of 71.5. The base is 7.435 and the high-finding penalty is 0.25. The final half-up result is 7.2.

Alongside the 7.2 score, the report shows all five dimension values and any ceiling or penalty. It records scope, priority actions, and the effect of a retest. The evidence set and calculation are included so the number can be reproduced.

6. Reporting and remediation

The customer package separates audiences while preserving one record:

• Executive view: scope and score; the dimension profile and material attack paths; business impact and priority decisions.
• Technical register: affected assets and evidence; reproduction and severity where appropriate; risk factors, remediation, owner, and the verification requirement.
• Coverage statement: sources used, inaccessible areas, exclusions, stale evidence, and assumptions.
• Remediation plan: immediate containment, near-term correction, architectural work, accepted risk, and target dates.
• Retest statement: what will be repeated, what constitutes closure, and how the score and risk delta will be recalculated.

A score is dated. Changes after the assessment do not silently rewrite it. A verified reassessment creates a new result and retains the previous model inputs for comparison.

7. Interpretation and limitations

The KernelHawk Baseline score is not a certification, warranty, breach prediction, compliance attestation, or substitute for a framework-specific audit. It describes the authorized scope and available evidence at a point in time. Two organizations with different scopes should not compare the number without the dimension and coverage context.

Equal weighting is intentionally simple and reproducible in model version 3. It may not match every organization's risk appetite. Professional judgment remains in selecting an observation state; the evidence does not automatically prove that selection correct. The result is defended by the closed anchors, recorded rationale, exact references, complete rubric coverage; finding reconciliation, canonical derivation, and evidence revision. Any future formula change requires a new version, published method, migration rule, and side-by-side impact analysis. A score produced with an earlier model must retain that model's label; it cannot be relabeled as version 3.

The criterion crosswalk to NIST CSF 2.0 and CISA Cross-Sector Cybersecurity Performance Goals is informative. It carries no endorsement or certification, and is neither an audit result nor a CSF Profile. Framework identifiers never contribute to score arithmetic.

Conclusion

The 1-10 number is a communication layer over an evidence record. Take away the authorisation gate, the published anchors or the verified follow-up and what is left is false precision with a decimal point.

Edition history: September 1, 2026, version 1.3. September 8, 2026, version 1.4: source links verified, copy simplified, and scoring explanation retained.