RepoJournal
AI Red Teaming AI Red Teaming
16 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 2026-10-07
stories 40

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

PyRIT grows an attack registry, promptfoo tightens JS assertion grading

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

The day's substantive work lands on the injection and grading paths operators rely on: PyRIT is building a class-only attack registry and indexed attack analytics, while promptfoo closes a run of JavaScript assertion grading bugs that could silently change your eval results.

fix(assertions): preserve async expression grading promptfoo/promptfoo

by mldangelo-oai

Declaration-prefixed JavaScript assertions could fail when async code used a regular expression after await, because statement boundaries were parsed against the wrong grammar. Grading now validates the same return-inserted body that actually executes, so bare grading objects, line separators, scores, and metadata survive intact; 479 affected tests pass.

fix(assertions): preserve keyword-property grading promptfoo/promptfoo

by mldangelo-oai

A companion fix covers declaration-based assertions where division, multiplication, or calls follow keyword-named properties. Acorn's property tokens are normalized while the original assertion text is preserved, keeping boolean, numeric, and GradingResult outcomes unchanged.

fix: handle assertion comments and dataset limits promptfoo/promptfoo

by mldangelo-oai

Inline assertions no longer mistake comment semicolons for statement boundaries, and malformed Hugging Face limits are validated before they reach the rows API, with explicit argument precedence and zero-request limits preserved. npm override traversal for known-bad resolution ranges was extended to scoped and version-qualified selectors.

fix: preserve Cloud errors and check npm aliases promptfoo/promptfoo

by mldangelo-oai

Cloud permission-denied responses with missing or malformed error data now preserve ConfigPermissionError instead of degrading into a generic failure, and non-403 responses still continue. Known-bad npm release checks now resolve dependency aliases to their target package and range first.

FEAT: Indexed attack analytics microsoft/PyRIT

by romanlutz

PyRIT adds indexed attack analytics, alongside a class-only attack registry foundation and nested and multiple conversion marker support, and fixes that pick the latest PyPI release and clean up SQLite on cancellation. garak, meanwhile, now respects parallel_capable when issuing generator requests.

Quick answers

What shipped in AI Red Teaming on October 7, 2026?
The day's substantive work lands on the injection and grading paths operators rely on: PyRIT is building a class-only attack registry and indexed attack analytics, while promptfoo closes a run of JavaScript assertion grading bugs that could silently change your eval results. In total, 21 commits and 19 pull requests landed.
Who contributed to AI Red Teaming on October 7, 2026?
5 developers shipped this update, including mldangelo-oai, mldangelo, richlundeen, romanlutz, and Denel John.
What were the notable AI Red Teaming updates?
fix(assertions): preserve async expression grading, fix(assertions): preserve keyword-property grading, and fix: handle assertion comments and dataset limits.