$ the-wire · showcase
Claude 5.2 lands in promptfoo, PyRIT bounds its tool loop
By RepoJournal · Filed · About AI Red Teaming · Composed from the cited sources · methodology
promptfoo adds Claude Fable 5.2 and Mythos 5.2 across four providers while fixing an Excel path bug that silently returned zero rows; PyRIT caps OpenAIResponseTarget's tool loop and stops a failing tool from dragging the turn on.
feat(providers): add Claude Fable and Mythos 5.2 promptfoo/promptfoo
promptfoo registers claude-fable-5-2 and claude-mythos-5-2 on the Anthropic API, Amazon Bedrock, Google Vertex AI and Azure AI Foundry, with request rules matching 5.1: thinking is always on, so disabled and budget_tokens are rewritten, and temperature and top_p are stripped. Most 5.1 references in docs, examples and UI pickers now point at 5.2.
fix(config): preserve literal hashes in Excel workbook filenames promptfoo/promptfoo
A local eval pointing at file://cases#2.xlsx#Data#2026 used to open cases, return no rows, or watch the wrong path; the reader now opens cases#2.xlsx, selects the worksheet Data#2026, and watches the physical workbook. The same rule covers a hash in the parent directory, so datasets with # in their filenames behave the same locally and in watch mode.
FIX: Strip judge-supplied original_float_value from parsed score metadata microsoft/PyRIT
normalize_score_to_float preferred a judge-supplied metadata.original_float_value over the validated score_value, applying an unvalidated float() with no [0,1] clamp, so a judge body of score_value "0.0" with original_float_value "1.0" scored as a pass. Raw float-scale scorers no longer let the judge's own key override the parsed score; FloatScaleThresholdScorer was unaffected because it writes...
FEAT: Bound OpenAIResponseTarget's tool loop, and end the turn when a tool's infrastructure fails (part 1) microsoft/PyRIT
Part 1 of 5 toward sandboxed code execution makes OpenAIResponseTarget's tool loop bounded, so a run can be left going instead of spinning indefinitely, and lets a tool end the turn when its own infrastructure fails. It needs no sandbox and stands alone, ahead of the four stacked PRs that follow.
FIX: Preserve unsent chat drafts per conversation across application navigation (#3735) microsoft/PyRIT
Elsewhere in PyRIT, unsent composer input is now preserved per conversation rather than in a single ChatDraftProvider slot, so navigating to History, Operations, Registry or Home no longer silently overwrites or discards a draft when you switch chats.