Sample assessment report

Ledgerline — what a report looks like

Ledgerline is a fictional multi-tenant invoicing product: our own demo application, built with deliberate weaknesses so the whole method can be shown end to end without touching anyone's real system. Everything below is real output from an assessment run on 3 October 2026, condensed into the format you would receive. Nothing here is a customer finding.

Confirmed live against the running application

Dependency fixes prepared, re-scanned clean, tests still green

804

Raw engine results condensed into this report

Confirmed

Reproduced against the running application

Each of these was proven with real requests against a deployed copy of the application, using two real accounts where the issue is about who can see what. They lead the report because nothing about them is a guess.

Confirmed Critical OWASP A02

Unsigned login tokens are accepted

JWT ‘alg: none’ · GET /api/auth/me

Proof

A real token from a real login was copied with its signature removed and its algorithm set to none. The protected endpoint returned 200 for both the genuine and the forged token.

Impact

The server does not check token signatures, so an attacker who edits the claims inside a token can act as another user.

Remediation

Pin the accepted algorithm when verifying tokens. Not auto-patched: authentication code is handed to your team with the exact failing request attached.

Confirmed Critical OWASP A02

Token signing secret is guessable

JWT HS256 weak secret · GET /api/auth/me

Proof

The signing secret was recovered offline on the first guess from a list of common weak secrets (six characters, shown masked as s****t). A fresh token signed with it was accepted live: 200.

Impact

Even with the algorithm pinned, an attacker can sign valid tokens.

Remediation

Rotate to a long random secret held in your secret store; invalidate existing sessions.

Confirmed High OWASP API1

One customer can read another customer's invoices

Broken object-level authorisation · GET /api/invoices/{id}

Proof

Two real accounts in two different organisations. User B requested invoice 4, which belongs to User A's organisation, and received it, including a marker value unique to that invoice.

Impact

Cross-tenant data exposure by changing a number in the URL.

Remediation

Scope the invoice lookup to the caller's organisation. Authorisation logic is left to a reviewer who knows your tenancy model.

Confirmed High OWASP LLM07

The in-app AI assistant reveals its system prompt

System prompt leakage · POST /api/copilot/chat

Proof

A known canary string planted in the system prompt, LEDGERLINE-CANARY-SYSTEM-PROMPT-9d41, came back verbatim, along with an internal “admin override phrase”. A plain baseline question returned a normal answer.

Impact

Internal instructions and anything embedded in them are readable by any user.

Confirmed High OWASP LLM01

The AI assistant follows injected instructions

Prompt injection · POST /api/copilot/chat

Proof

An instruction to override its guidelines and repeat a unique marker was followed; the marker prompt-injection-verify-6d2bf37cb07cbf39 came back verbatim, twice.

Impact

User input can redirect the assistant, including toward the tools it can call.

In this demo the assistant is a deterministic stand-in for a model, so results are reproducible. Against a real model the same checks run unchanged.

Corroborated

Two independent engines agree

Found by static analysis and flagged independently by two engines from different vendors. Strong signal, but not exercised live, and labelled that way.

Critical & High Dependencies · fix verified

Known-vulnerable packages: lodash, minimist, node-fetch

package-lock.json · CVE-2021-44906, CVE-2021-23337, CVE-2020-8203, CVE-2022-0235 and others

Agreement

Reported by both dependency engines.

Fix prepared

Each moved to the lowest version that resolves it: lodash 4.17.15 → 4.18.0, minimist 1.2.5 → 1.2.6, node-fetch 2.6.0 → 2.6.7.

Re-scan

All three confirmed fixed by a re-scan of the patched tree, and independently by a second engine.

Tests

The application's own test suite: 19 of 19 passing before the change and 19 of 19 after.

High Dependencies · no fix prepared

Web framework with published bypass advisories: fastify 4.29.1

package-lock.json · authentication and request-validation bypass advisories

Agreement

Reported by both dependency engines.

Status

Upgrade not prepared in this pass. Reported as open rather than quietly left out.

Critical Source code

Hard-coded secret in application config

src/config.ts:8

Agreement

Flagged independently by two static analysis engines.

Critical & High Infrastructure

Cluster and cloud configuration

deploy/k8s/ledgerline.yaml · deploy/terraform/main.tf · Dockerfile

Kubernetes

A cluster role with every verb on every resource (line 82); the app container runs privileged with SYS_ADMIN (line 33).

Cloud

A firewall rule open to any address (line 69); a storage bucket that allows public access (line 95).

Container

The image runs as root. Its end-of-life base image carries 512 known-vulnerable packages; a trial rebuild on a current slim base built cleanly and cut that to 330 (critical 14 → 9, high 151 → 88).

Agreement

Each configuration item flagged by both infrastructure engines.

Single engine

Worth a look, labelled as unproven

One engine saw these and nothing else confirmed them. They are reported, ranked last, and never presented as vulnerabilities.

Single engine High

SQL string assembled from request input

src/routes/users.ts:41

Status

Pattern match from one engine; the second did not flag it and it was not reproduced live. Needs a reviewer, not a fire drill.

Limits

What this assessment did not cover

An unauthenticated passive web scan plus API schema fuzzing found one low-severity item (cacheable responses). The serious issues above are about authorisation, sessions and the AI assistant, which that kind of scan cannot see; that is why the targeted, logged-in checks exist.
Infrastructure fixes were not prepared in this run. They are model-assisted and that option was switched off; the findings stand without them.
No live cloud account was in scope, so deployed cloud configuration was not assessed — only the Terraform that describes it.
Of the 804 raw results, 522 are individual vulnerable packages inside the container image. The full list ships as a machine-readable appendix (JSON and SARIF) rather than as pages of this report.

Want this for your own application?

The first assessment is on us: one repository, one report, no obligation.

Request an assessment Reply within two business days.