$ the-wire · showcase
PyRIT adds a fullwidth text converter, docs its audio output config
By RepoJournal · Filed · About AI Red Teaming · Composed from the cited sources · methodology
PyRIT shipped one new attack-side capability operators can adopt today, a fullwidth-character converter for probing unnormalized filters, and filled the two biggest documentation gaps around its data-driven and audio targets; promptfoo's day was dependency bumps only.
FEAT: Add FullwidthConverter microsoft/PyRIT
FullwidthConverter in pyrit/converter/fullwidth_converter.py maps printable ASCII to Unicode fullwidth forms (Hello! -> Hello!), which normalizes back to the original under NFKC. That makes it a ready probe for filters or classifiers that match exact characters without normalizing input first; word selection is configurable and spaces between words and non-ASCII characters pass through unchanged.
MAINT: declare LatentInjection ingredients as required data microsoft/PyRIT
The five LatentInjection ingredient datasets now go through get_required_data and the reserved boundary marker check became a required data validator, so refreshed data gets validated too. LatentInjectionDatasetConfiguration drops its own uncapped and exact-datasets checks while keeping the families, max_total and family/trigger sampling; filters and fetch settings on the config still reach the...
DOC: Document OpenAIChatAudioConfig audio output for OpenAIChatTarget microsoft/PyRIT
The OpenAI Chat Target page now has an Audio Output section covering voice, audio_format and prefer_transcript_for_history on OpenAIChatAudioConfig, how it merges into extra_body_parameters, and what an audio reply returns: a transcript text piece and a saved audio_path piece. The docs also note audio output is Chat Completions only, pointing to the realtime target for the other path.
TEST: Add opt-in GPU smoke test for GCG on a real checkpoint microsoft/PyRIT
test_gcg_gpu_smoke_integration.py runs GCGGenerator for five steps against the full Qwen2.5-0.5B-Instruct checkpoint on a CUDA GPU with real worker subprocesses, checking loss goes down and that the final suffix's recomputed target loss matches the logged one. It is opt-in, so it costs nothing in the default suite.
MAINT: Add static keyword-only parameter check with baseline microsoft/PyRIT
A static keyword-only parameter check now covers the rest of pyrit/ beyond the pluggable base-class constructors, landed against a baseline so no signature changes today. Inert for operators, but it is the mechanism behind future signature changes.
chore(deps): update dependency @babel/core to v8 promptfoo/promptfoo
Also landed: promptfoo bumped tsdown to 0.23.4 and @anthropic-ai/claude-agent-sdk to 0.3.311, and moved @babel/core from ^7.29.6 to ^8.0.0; nothing in these touches promptfoo eval or red-team behavior.