Skills
Skills
What a skill is, the built-in catalogue by category, when a skill runs, and how to switch skills on and off for the organisation or for one repository.
A skill is a packaged capability that Nullgap AI can use: what it is for, when it applies, the playbook it follows, the tools it may call and what it returns. Nullgap ships a catalogue of built-in skills and uses the ones that apply during pull request reviews, and during pentests when an admin has switched that on. A skill that allows it can also be run by itself against a repository, a pull request, a finding or two versions of an artefact.
Your organisation can switch skills on and off, override that for one repository, and add skills of its own.
A skill is not a custom check. The difference, and which one to use, is on Custom checks or skills.
#Where to find them
Open Skills in the console. The page has three tabs:
| Tab | What it shows |
|---|---|
| Catalogue | Every skill, built-in and custom, grouped by category, with its switch |
| Custom skills | The skills your organisation wrote or uploaded |
| Run a skill | Starting a run on demand, and the recent runs |
Every member can read the catalogue and start a run. Admins and owners change switches and manage custom skills.
Through the API:
curl -s -H "Authorization: Bearer $NULLGAP_API_KEY" \
"$NULLGAP_URL/api/orgs/$NULLGAP_ORG/skills"
The answer has skills (one entry each, with its manifest, whether it is enabled and the state of each tool it names), tools, your settings, the categories and the limits for custom skills. One skill with its full playbook is GET /api/orgs/:org/skills/:id.
#When a skill runs
Each skill says where it may run:
| Where | Meaning |
|---|---|
| Pull request reviews | Runs inside a review when it applies to the changed files |
| Pentests | Runs inside a pentest when it applies to the repository. Built-in skills only when Run built-in skills in pentests is on; your custom skills always |
| Live testing | Joins a pentest only when that pentest has a live-testing target |
| On demand | Can be started by a person, an API key or an MCP client |
A skill applies to a repository according to its triggers:
- File patterns and technologies are triggers. When a skill names either, at least one must be present in the repository; in a review, among the changed files.
- Languages narrow further: when a skill names any, the repository must contain code in one of them.
- A skill that names none of these applies to every repository.
One review uses at most 4 skills and one pentest at most 12. Skills triggered by files that are actually present go first. A skill that applied but was left out, for the limit or because it is switched off, is recorded with the reason, so the result can say what did not run.
Only skills that return findings take part in reviews and pentests. A skill that returns a document or a triage verdict runs on demand only.
#Built-in skills in pentests
The built-in skills that return findings do not run inside pentests unless an admin turns on Run built-in skills in pentests. In real pentests they mostly repeated what the main analysis and the scanning tools had already found, and each one makes the pentest take longer. Your own custom skills run in pentests either way, finding-triage still guides the independent check, and pull request reviews are not affected. Each built-in skill that was left out for this reason is recorded in the pentest's skill statistics with the reason built-in skills are off for pentests.
In the console, open Skills: the switch is at the top of the Catalogue tab. Through the API, with an admin key:
curl -s -X PUT -H "Authorization: Bearer $NULLGAP_API_KEY" \
-H "Content-Type: application/json" -d '{"builtinInPentests":true}' \
"$NULLGAP_URL/api/orgs/$NULLGAP_ORG/skills/settings"
Each change is written to the audit log as skills.builtin_in_pentests_on or skills.builtin_in_pentests_off.
#What a skill adds to a review or a pentest
What a skill raises during a review or a pentest is a candidate like any other. It is merged with duplicates, goes through the same independent check, and serious ones are reproduced in the sandbox. The finding carries the skill's name: finding.source is llm-skill: followed by the skill's id, and the finding's trace names it.
A skill influences what is looked at. It does not decide the form of the result or the assurance a finding carries: no skill can mark anything verified or proven. Those assurance levels are set by the independent check and by a reproduction, and by nothing else.
#The built-in catalogue
There are 33 built-in skills in six categories. The Returns column says whether a skill produces findings, a document or a triage verdict; Runs in says where.
#Application security
| Skill id | What it covers | Runs in | Returns |
|---|---|---|---|
secure-code-review | A general security review of a change or a repository | Reviews, on demand | Findings |
injection-and-input-handling | Injection flaws and unsafe handling of input | Pentests, on demand | Findings |
authentication-and-session-review | Sign-in, sessions and credential handling | Pentests, on demand | Findings |
authorization-and-tenant-isolation | Access checks and separation between tenants | Pentests, on demand | Findings |
api-security | HTTP APIs and their descriptions | Pentests, on demand | Findings |
business-logic-abuse | Flows that can be used in ways the design did not intend | Pentests, on demand | Findings |
cryptography-and-secrets-in-code | Misused cryptography and secrets written into code | Pentests, on demand | Findings |
deserialization | Unsafe deserialisation of untrusted data | Pentests, on demand | Findings |
file-and-parser-handling | Uploads, archives, paths and parsers | Pentests, on demand | Findings |
ssrf-and-outbound-requests | Requests the server makes on a caller's behalf | Pentests, on demand | Findings |
memory-safety | Memory-safety flaws in native code | Pentests, on demand | Findings |
llm-application-security | Applications that call a language model | Pentests, on demand | Findings |
sast-triage | Sorting static-analysis results into real and not real | Pentests, on demand | Findings |
dast | Dynamic testing of a running application | Live testing only | Findings |
finding-triage | The playbook of the independent check, on one finding | On demand | A triage verdict |
#DevSecOps
| Skill id | What it covers | Runs in | Returns |
|---|---|---|---|
cicd-pipeline-review | Build and deployment pipelines | Reviews, pentests, on demand | Findings |
infrastructure-as-code | Infrastructure definitions and cluster manifests | Reviews, pentests, on demand | Findings |
container-image-and-dockerfile | Container builds and images | Reviews, pentests, on demand | Findings |
cloud-identity-and-permissions | Cloud roles and permissions in code | Reviews, pentests, on demand | Findings |
software-supply-chain | Dependencies and how they are obtained | Reviews, pentests, on demand | Findings |
#Secret detection
| Skill id | What it covers | Runs in | Returns |
|---|---|---|---|
secret-detection | Credentials in the code and in its history | Reviews, pentests, on demand | Findings |
#Secure SDLC
| Skill id | What it covers | Runs in | Returns |
|---|---|---|---|
threat-modeling | A threat model of the system | On demand | Document: threat model |
security-requirements | Testable security requirements for a feature | On demand | Document: security requirements |
secure-design-review | A review of a design before it is built | On demand | Document: design review |
security-test-ideas | Security test cases for a feature or a change | On demand | Document: test plan |
#Compliance
| Skill id | What it covers | Runs in | Returns |
|---|---|---|---|
compliance-mapping | Which controls of a framework the evidence relates to | On demand | Document: control mapping |
policy-enforcement | Whether the repository follows a policy you state | On demand | Document: policy evaluation |
audit-evidence-pack | Evidence collected and indexed for a review by others | On demand | Document: evidence pack |
These three produce working documents for your own use. Each states what was examined and what cannot be concluded. None of them is an attestation, and none makes a statement about your organisation's standing against a framework.
#Incident response
| Skill id | What it covers | Runs in | Returns |
|---|---|---|---|
leaked-credential-response | What to do about an exposed credential | On demand | Document: incident runbook |
malicious-dependency-response | What to do about a compromised package | On demand | Document: incident runbook |
exploited-vulnerability-response | What to do about a finding that is being exploited | On demand | Document: incident runbook |
security-event-analysis | A timeline from logs or an alert you supply | On demand | Document: incident timeline |
patch-diff-analysis | What a patch fixed, from two versions of an artefact | On demand | Document: patch analysis |
The catalogue in your console is the authority: it lists exactly what your installation ships, with each skill's full description, inputs, tools and limits.
#Tools
A skill names tools by id, never a command. A tool is one of the built-in ones (the code index, the static rules, dependency data, git history, the sandbox, live requests) or a scanner installed on the Nullgap installation.
For each tool a skill names, the catalogue shows one of three states:
| State | Meaning |
|---|---|
| Available and permitted | The skill can call it |
| Not available | It is not installed on this installation. The skill still runs, and the run records that the tool did not run |
| Not permitted | Your organisation has restricted the tools skills may use, and this one is not on the list |
An unavailable tool never means "nothing found": the output says the tool did not run.
Tools that execute anything run in the sandbox. Tools that send requests work only inside a pentest that has a live-testing target, and every request stays inside the scope list rules of that pentest.
Admins can restrict which tools skills may call with allowedTools in the skill settings: null permits every tool the installation offers, a list permits only those.
#Switching a skill on or off
In the console, open Skills, find the skill on the Catalogue tab and use its switch. This is for admins and owners, and each change is written to the audit log as skill.enabled or skill.disabled.
Through the API, with an admin key:
curl -s -X POST -H "Authorization: Bearer $NULLGAP_API_KEY" \
-H "Content-Type: application/json" -d '{}' \
"$NULLGAP_URL/api/orgs/$NULLGAP_ORG/skills/secret-detection/disable"
POST /api/orgs/:org/skills/:id/enable switches it back on. Both answer with the new settings:
{
"disabled": ["secret-detection"],
"repositories": {},
"allowedTools": null,
"builtinInPentests": false
}
A skill that is switched off takes no part in reviews and pentests and cannot be run on demand: starting it answers 400 skill_disabled.
finding-triage cannot be switched off. It is the playbook of the independent check that every finding goes through, and trying answers 400 always_on.
#Overrides for one repository
A repository can differ from the organisation: a skill that is off for the organisation can be forced on for one repository, and a skill that is on can be forced off there.
In the console, on the Catalogue tab, choose the repository under Show and change settings for, then set a skill to Force on or Force off.
Through the API, send the repository's id with the same two calls:
curl -s -X POST -H "Authorization: Bearer $NULLGAP_API_KEY" \
-H "Content-Type: application/json" \
-d '{"repoId":"repo_4c1d2e3f5a6b"}' \
"$NULLGAP_URL/api/orgs/$NULLGAP_ORG/skills/infrastructure-as-code/disable"
An override is kept only where it differs from the organisation's own switch: enabling a skill for a repository where the organisation already has it on removes the override instead of recording one. The settings then look like this:
{
"disabled": [],
"repositories": {
"repo_4c1d2e3f5a6b": { "enabled": [], "disabled": ["infrastructure-as-code"] }
},
"allowedTools": null,
"builtinInPentests": false
}
To set all overrides and the permitted tools in one call, use PUT /api/orgs/:org/skills/settings with repositories (a map of repository id to { "enabled": [...], "disabled": [...] }) and allowedTools. The map you send replaces the stored one.
#From the command line and from an agent
- The CLI lists and switches skills:
nullgap skills list,nullgap skills enable <name>andnullgap skills disable <name>. They readNULLGAP_URL,NULLGAP_ORGandNULLGAP_ORG_API_KEY; switching needs a key with theadminscope. - An MCP client can read the catalogue with
list_skills, start a run withrun_skilland read it withget_skill_run. See MCP tools.
Nullgap documentation