RepoJournal
AI Red Teaming AI Red Teaming
52 wires and counting

$ follow AI Red Teaming

Keep up with AI Red Teaming in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-11-18
stories 58

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

PyRIT hardens converters and scoring, drops write-only deliveries from results

By RepoJournal · Filed · About AI Red Teaming · Composed from the cited sources · methodology

PyRIT shipped a batch of correctness fixes that change how results are scored, how converters build prompts, and what conversation storage accepts, several of which will alter existing scan output or raise errors at construction.

[BREAKING] FIX: Do not score write-only target deliveries as attack responses (#3757) microsoft/PyRIT

by Dev3 ^~^

Deliveries to write-only targets are no longer recorded as attack responses, so scans against targets that do not return content will no longer log fabricated successes in your result set. The fix is flagged breaking, which means previous scores from write-only targets are not comparable to what you get now.

FIX: Require ordered template parameters in LLMGenericTextConverter microsoft/PyRIT

by samrudhi03

LLMGenericTextConverter rejected set and frozenset template parameters at construction with a ValueError, because their iteration order varies with PYTHONHASHSEED and the same config could emit different prompts across processes. If any of your converter configs pass sets, they will now fail to build until you switch to an ordered input.

FIX: Reject non-dictionary metadata before conversation updates microsoft/PyRIT

by YingzuoLiu

Non-dictionary prompt_metadata roots now raise ValueError before anything is written to storage, instead of breaking later message retrieval. A mixed conversation update carrying invalid metadata leaves all fields unchanged, so a rejected write no longer half-applies.

FIX Insert tatweel only between Arabic letters that join in TatweelConverter microsoft/PyRIT

by ShivamSharma43

TatweelConverter used to insert U+0640 between every pair of adjacent Arabic letters, producing detached strokes in words like the documented example output. It now inserts tatweel only where the preceding letter actually joins, using the 52 right-joining and 2 non-joining letters in the U+0600 to U+06FF range, so tatweel-based perturbation prompts render differently than before.

FIX: Bind Attack History requests to the runtime generation microsoft/PyRIT

by itsrayeedscodespace

Attack History requests, pagination and filters are now bound to the current runtime generation: late responses from an obsolete runtime are ignored, and switching runtimes resets pagination without carrying over the old cursor, filter options or errors.

Quick answers

What shipped in AI Red Teaming on November 18, 2026?
PyRIT shipped a batch of correctness fixes that change how results are scored, how converters build prompts, and what conversation storage accepts, several of which will alter existing scan output or raise errors at construction. In total, 29 commits and 29 pull requests landed.
Who contributed to AI Red Teaming on November 18, 2026?
6 developers shipped this update, including ianw-oai, Dev3 ^~^, itsrayeedscodespace, YingzuoLiu, samrudhi03, and ShivamSharma43.
What were the notable AI Red Teaming updates?
[BREAKING] FIX: Do not score write-only target deliveries as attack responses (#3757), FIX: Require ordered template parameters in LLMGenericTextConverter, and FIX: Reject non-dictionary metadata before conversation updates.