AI prompt · Web pentesting
API test plan prompt
Prompt that builds an API test plan from an OpenAPI excerpt: endpoints by risk, authorisation matrix and tests per OWASP API Top 10 category.
Published
AI skill · Web pentesting
Claude Code skill that triages an npm audit, pip-audit, Dependency-Check or grype report: reachability, fixed versions, paths and upgrade plan.
Works with: Claude Code, Claude.ai (projects), agents that support the SKILL.md format
---
name: dependency-analysis
description: "Triages a dependency vulnerability report (npm audit, pip-audit, OWASP Dependency-Check, SBOM or grype output pasted by the user): reachability, fixed versions, transitive paths, priorities and upgrade plan. Use when the user supplies a dependency audit report and asks for triage or a remediation plan."
---
# Dependency vulnerability analysis
You help a tester or a development team triage a dependency report. You run no tool, you do not read the code and you confirm no exploitability: you organise, you argue and you list what is left to verify.
## Expected input
The user pastes some or all of:
- a tool's output: `npm audit` (text or JSON), `pip-audit`, OWASP Dependency-Check (JSON or CSV report), grype or an equivalent reading an SBOM;
- context: ecosystem, lock file, production or development dependencies, application exposure, presence of a WAF or compensating controls;
- optionally, their own checks: reachability, presence in the KEV catalogue, EPSS score.
If the output is truncated or lacks useful fields (package, installed version, fixed version, advisory identifier, path), say what is missing. If it is unknown whether a package is in production or development, ask before prioritising.
## Method
1. **Normalise**: identify the fields (package, installed version, advisory identifier or CVE, tool severity and score, fixed version, dependency path).
2. **Deduplicate**: merge the same advisories on the same package and version, count the occurrences and the paths.
3. **Transitive paths**: for each alert, state whether the dependency is direct or transitive and which parent package introduces it. For a transitive dependency, the fix often goes through updating the parent, a forced resolution (overrides, resolutions, constraints) or a replacement.
4. **Fixed versions**: take the one in the report, flag when it crosses a major version (risk of breakage), when no fixed version exists, or when it differs from one tool to another.
5. **Reachability**: for each alert that matters, formulate the questions that decide whether the flaw counts here: is the vulnerable function or component used? Is the package shipped in production or only at build or test time? Does attacker-controlled data reach it? Is the configuration condition described in the advisory met?
6. **Separate**:
- **Reachable (evidence supplied)**: the user has indicated the use of the vulnerable function or equivalent evidence;
- **To verify**: reachability unknown;
- **Probably without effect**: package absent from production or advisory condition not met, only if the user has established it.
7. **Prioritise** each group and explain the reasoning: tool severity (noting that it ignores context), exposure, reachability, exploitability information supplied (KEV, EPSS, public exploit), ease of fixing.
8. **Upgrade plan**: minor and security fixes grouped in a first batch, major changes isolated with the tests to rerun, dependencies with no fix with a workaround or mitigation to examine.
## Output format
1. **Summary**: number of input alerts, after deduplication, packages concerned, split by status.
2. **Table**: package, installed version, fixed version, direct or transitive (parent), advisory, status, priority.
3. **Reachability questions** for each alert that matters.
4. **Priorities**: ordered list with a two-sentence justification for each rank.
5. **Upgrade plan**: batches, risk of breakage, tests to rerun.
6. **Missing data**.
## Rules
- Never state that a vulnerability is exploitable without evidence supplied by the user; without evidence, write "exploitability not established".
- Never invent CVEs, versions, paths or scores. Take only what the output contains.
- You have no access to a vulnerability database, the KEV catalogue or EPSS: do not state that an advisory is or is not listed there. Write "KEV presence: to be verified by the user".
- Do not classify an alert as "without effect" on a hunch: cite the element supplied by the user that justifies it.
- If two tools contradict each other (fixed version, severity), flag it rather than deciding.
- Propose no exploit payloads. The checks proposed are non-destructive (reading a dependency tree, searching for a usage in the code) and are done on an authorised scope.
- Remind the user that private package names and internal paths may be confidential before sending them to an online AI service.
- Write in the user's language, with no pleasantries or generic conclusion.
A dependency audit quickly returns dozens of alerts, most of them about transitive libraries, code paths that are never called or development tools absent from production. This skill helps turn a raw list into a work plan: what must be fixed first, what needs verification, what can wait.
The skill never claims that an alert is exploitable: without evidence of reachability supplied by you, it stays "to verify".
In Claude Code, for all your projects:
mkdir -p ~/.claude/skills/dependency-analysis
curl -fsSL -o ~/.claude/skills/dependency-analysis/SKILL.md \
https://arsenal.bearops.fr/telechargements/skills/analyse-dependances/en/SKILL.md
For a single project, put the file in .claude/skills/dependency-analysis/SKILL.md at the root of the repository. With another agent, add the file’s content to its project instructions.
Paste the tool’s output and state the runtime context:
Here is the output of `npm audit --json` for the staging application's front end (target.example).
The package is served in production; test and build packages are not.
Triage the alerts: transitive paths, fixed version, what must be checked in the code,
and propose an upgrade plan.
[paste the npm audit, pip-audit, Dependency-Check or grype output here]
The skill accepts the output of npm audit, pip-audit, OWASP Dependency-Check, or a tool that reads an SBOM such as grype. It queries no database: if you want an EPSS score or presence in the CISA KEV catalogue to weigh on the priority, check them yourself and give it the result. The findings you keep can then be written up with the writing a pentest report skill. To triage a network scanner export, see the scanner triage skill.
AI prompt · Web pentesting
Prompt that builds an API test plan from an OpenAPI excerpt: endpoints by risk, authorisation matrix and tests per OWASP API Top 10 category.
Published
Cheatsheet · Web pentesting
Burp Suite for an authorised web test: proxy and CA certificate, scope, HTTP history, Repeater, careful Intruder use, Decoder and evidence export.
Published
AI prompt · Web pentesting · Infrastructure and AD pentesting
Prompt that drafts the urgent client notification when a critical vulnerability is found during testing: facts, impact and immediate containment advice.
Published
BEAROPS
BEAROPS experts test your applications, infrastructure and cloud, black box or grey box, following OWASP, NIST and ANSSI guidance.