> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://nemo-platform.docs.buildwithfern.com/nemo/platform/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://nemo-platform.docs.buildwithfern.com/nemo/platform/_mcp/server.

# Selecting Probes

<a id="auditor-configs-probes" />

An `AuditConfig`'s `plugins.probe_spec` field controls which garak probes run during an audit. Two related fields refine the selection: `plugins.detector_spec` chooses how outputs are scored, and `run.probe_tags` filters the selected probes by hazard tag.

## `probe_spec`

`probe_spec` is a comma-separated string. Each entry is one of:

* `"all"` — every probe in the garak library.
* A category — every probe in that category (for example, `latentinjection` runs every `latentinjection.*` probe).
* A fully-qualified probe class — `category.ProbeClass` (for example, `encoding.InjectROT13`).

Combine entries with commas. There is no negation syntax; if you need a subset of a category, list the probes explicitly.

### Examples

A single category, useful when you want broad coverage of one failure mode:

```python
AuditPluginsData(probe_spec="latentinjection")
```

Two specific probe classes — useful for fast smoke tests:

```python
AuditPluginsData(probe_spec="encoding.InjectROT13,dan.AutoDANCached")
```

A category mixed with a specific class from another category:

```python
AuditPluginsData(probe_spec="latentinjection,goodside.Tag")
```

## `probe_tags`

`run.probe_tags` is a comma-separated list of tags (such as `"owasp:llm06"` or `"payload:hallucination"`). When set together with `probe_spec`, the probes selected by `probe_spec` are then **filtered to only those that also match the tag list** — the two fields combine by intersection, not union.

```python
AuditConfig(
    name="owasp-llm06-scan",
    workspace="default",
    plugins=AuditPluginsData(probe_spec="all"),
    run=AuditRunData(probe_tags="owasp:llm06"),
)
```

Leave `probe_tags` unset (the default `None`) to run every probe selected by `probe_spec` without tag filtering.

## `detector_spec`

`plugins.detector_spec` selects how garak scores each generated response. The default `"auto"` lets each probe choose its own detectors — appropriate for most audits.

```python
AuditPluginsData(probe_spec="latentinjection", detector_spec="auto")
```

Specify a detector explicitly only when you have a reason to override the probe-recommended default:

```python
AuditPluginsData(
    probe_spec="latentinjection",
    detector_spec="dan.DAN",
)
```

## Where to Find the Full Probe List

The probe library is owned by the [garak project](https://github.com/NVIDIA/garak). For canonical, up-to-date documentation:

* [Probes reference](https://reference.garak.ai/en/latest/probes.html) — every probe, grouped by category, with descriptions.
* [Probe tiers](https://reference.garak.ai/en/latest/garak.probes._tier.html) — garak's importance tiers (1 is highest), useful for picking a small representative sample.
* [garak on GitHub](https://github.com/NVIDIA/garak) — source for the probe taxonomy and tags.

The plugin invokes the garak version installed in the interpreter at `~/.auditor/.venv/bin/python` (or wherever `NEMO_AUDITOR_GARAK_PYTHON` points). If a probe name is unknown, check that your installed garak version exposes it.