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 2026-10-12
stories 64

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

Promptfoo drops stale caches, PyRIT ships an attack analytics SDK

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

Promptfoo removed the config and MCP evaluation caches that hid fresh edits and new runs, while PyRIT added an attack analytics SDK and an Anthropic model-written evals loader operators can pull through SeedDatasetProvider.

fix: remove stale MCP evaluation summary cache promptfoo/promptfoo

by mldangelo-oai

MCP list_evaluations now reads summary rows on every request, so a long-lived server surfaces evaluations that were just created, updated, or deleted. The trade is explicit in the PR: repeated listings cost extra database reads, and the cacheStats fields come back zeroed.

fix(config): remove the default configuration cache promptfoo/promptfoo

by mldangelo-oai

Default-config discovery no longer retains parsed config objects or cached missing-file results, so declarative edits, new files, deletions, and parse failures all show up on the next load. CommonJS .js configs still work: successful VM fallback exports are cached for discovery only, and watch reruns invalidate them.

FEAT: Attack analytics SDK microsoft/PyRIT

by romanlutz

PyRIT merged an attack analytics SDK, giving operators programmatic analysis over attack results instead of ad hoc post-processing.

FEAT: Add Anthropic model-written evals dataset microsoft/PyRIT

by Steeve-Crypto

The Anthropic model-written evals land as a remote seed dataset loader in PyRIT, fetched through SeedDatasetProvider from the full anthropics/evals release rather than Hugging Face, which publishes only a subset. Supported collections are persona, sycophancy, advanced-ai-risk (human-written and LM-written), and winogenerated examples.

refactor(webui): consolidate media URL resolution promptfoo/promptfoo

by mldangelo-oai

The eval table now resolves storage and blob audio URLs through the shared media helpers, dropping the duplicate mediaStorage module, its loading state, and its effect; the audio player keys off the resolved URL so a source change reloads the right clip. The unreachable performance-history chart and its dead flag are also gone, with the three active charts unchanged.

Quick answers

What shipped in AI Red Teaming on October 12, 2026?
Promptfoo removed the config and MCP evaluation caches that hid fresh edits and new runs, while PyRIT added an attack analytics SDK and an Anthropic model-written evals loader operators can pull through SeedDatasetProvider. In total, 32 commits and 32 pull requests landed.
Who contributed to AI Red Teaming on October 12, 2026?
3 developers shipped this update, including mldangelo-oai, Steeve-Crypto, and romanlutz.
What were the notable AI Red Teaming updates?
fix: remove stale MCP evaluation summary cache, fix(config): remove the default configuration cache, and FEAT: Attack analytics SDK.