[FEAT] Add SemGuard Arabic Security Dataset loader

[Abdallah Abughallous (AbdaullahAG)]

Description

Adds a SeedDataset loader for the SemGuard Arabic Security
Dataset — 807 validated Arabic/Arabizi/English prompt attack
examples across 7 threat categories (injection/jailbreak,
phishing, privacy leakage, violent incitement, harmful content,
impersonation, safe). Validated via a three-judge LLM-as-Judge
pipeline (GPT-4o, Grok-4, Llama 3.3 70B; Fleiss' kappa = 0.839).

Closes #2331.

Implementation notes, per the discussion in #2331:

• Category enum: SemGuardCategory (7 values) lets callers
  filter by threat category, following the VLGuardSubset pattern.
  SAFE is included as an explicit filterable value rather than
  always loaded, per Roman Lutz (@romanlutz)'s point about excluding safe
  prompts by default.
• Language accessibility: each seed's metadata includes a
  language field (ar, en, or code-switched), so users can
  filter or at least identify non-English prompts programmatically
  even without reading Arabic. Category-level docstrings and the
  loader's own docstring describe each category in English. Open
  to adding English glosses per-seed if that would help further.
• harm_categories mapping: PyRIT's HarmCategory taxonomy
  doesn't have a direct entry for "prompt injection" as an attack
  technique (as opposed to a content-harm type), so
  injection_jailbreak maps to COORDINATION_HARM via
  HARM_CATEGORY_ALIAS_OVERRIDES. Flagging this in case a more
  precise mapping is preferred.
• The dataset itself is CC BY 4.0 licensed (separate from this
  loader file's MIT license, consistent with how gated/externally
  licensed datasets like VLGuard are handled).

Paper: accepted at IEEE AEECT 2026 (citation added to
references.bib / bibliography.md).

Tests and Documentation

• Added tests/unit/datasets/test_semguard_dataset.py: 9 tests
  covering happy-path fetch, category filtering, empty-after-filter
  raises ValueError, empty-text rows are skipped, dataset_name
  property, HARM_CATEGORY_ALIAS_OVERRIDES coverage for all 7
  categories, and SemGuardCategory enum values. All passing
  locally (pytest tests/unit/datasets/test_semguard_dataset.py -v
  → 9 passed).
• Registered the loader and SemGuardCategory in
  pyrit/datasets/seed_datasets/remote/__init__.py.
• Added BibTeX entry (abughallous2026semguard) to
  doc/references.bib and the cite key to doc/bibliography.md.
• Updated doc/code/datasets/1_loading_datasets.py (prose list)
  and regenerated the notebook via:
  jupytext --to ipynb --execute doc/code/datasets/1_loading_datasets.py
  — ran clean, confirmed semguard appears in the executed
  get_all_dataset_names_async() output cell alongside the other
  built-in datasets.

[Abdallah Abughallous (AbdaullahAG)]

@microsoft-github-policy-service agree