v0.4.0 — SecLists Integration & Wordlist Parsing

New Features

• SecLists LLM_Testing support — ingest prompts from SecLists Ethical and Safety Boundaries, Bias Testing, Data Leakage, Divergence Attacks, and Memory Recall categories out of the box
• Wordlist auto-detection — .txt files with 5+ lines and short median line length are automatically parsed as one-prompt-per-line wordlists instead of single prompts
• Placeholder expansion — prompts containing [GENDER], [COUNTRY], [REGION], [NATIONALITY], or [SKIN_COLOR] are expanded into concrete variants during parsing; geographic placeholders co-vary by index for consistency
• New technique classifiers — bias-probe, divergence, data-leakage, and policy-violation categories added to automatic prompt classification

Improvements

• CSV parser now recognizes question and content_policy_name columns (SecLists forbidden_question_set.csv compatibility)
• prompt-leak technique keywords expanded to catch system prompt extraction patterns (verbatim, print your instructions)
• README updated with Supported Datasets section documenting OBLITERATUS, SecLists, and custom dataset workflows
• 7 new tests (72 total) covering wordlist parsing, CSV question field, placeholder expansion, and new technique classification

Compatibility

• Validated against SecLists LLM_Testing: 659 prompts parsed from 5 files across all categories
• Existing OBLITERATUS parsing unchanged — long-form prompts continue to be treated as single entries

Full Changelog: v0.3.0...v0.4.0

v0.3.0 — Stakeholder Reporting & Judge Refinements

New Features

• Executive summary in HTML reports — color-coded risk badge, plain-English findings, and auto-generated remediation recommendations derived from actual detection results
• CVSS-AI score explainer — visual severity scale with links to CVSS v4.0, OWASP Top 10 for LLMs, and NIST CVSS Calculator
• darkarts report reproduce — generates markdown finding cards with exact prompts, classification, detection details, and step-by-step reproduction instructions a human tester can follow
• Automated PyPI publishing via GitHub Actions with trusted publisher (OIDC) and version tag verification

Improvements

• Judge rubric refined — harmful-content rubric now distinguishes between a model repeating/quoting adversarial content and actually complying with it (4-point anchored scale with 0.3 meta-analytical tier)
• Meta-analysis detector — code-level safety net using trigram overlap ratio and signal phrase matching to catch false positives where the model analyzes the prompt rather than executes it; high-confidence detections cap the judge score at 0.3
• README rewritten — full requirements walkthrough (Python, Git, Ollama setup), quick start guide, complete command reference, CVSS-AI scoring explanation

Fixes

• Removed stale PyPI badge and broken pip install darkarts instruction from README
• Fixed license classifier conflict with PEP 639 on Python 3.14
• Test matrix narrowed to 3.10–3.13 (stable releases)

Full Changelog: v0.2.0...v0.3.0

v0.2.0

Release notes here