$ cat ai-red-teaming/week/2026-09-28.log
the week in review · Sep 28 – Oct 4, 2026
Promptfoo makes Slack and WatsonX SDKs opt-in
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
Both are breaking changes to default installs: Slack needs @slack/web-api@^8.1.1, WatsonX its two IBM packages.
perf(deps)!: make Slack SDK opt-in promptfoo/promptfoo
Ordinary Promptfoo installs no longer pull the Slack SDK or the unused direct ANSI tokenizer dependency, and selecting Slack now validates the installed SDK and prints installation guidance when it is missing or an incompatible version. If your pipelines post results to Slack, add @slack/web-api@^8.1.1 alongside Promptfoo in the same environment.
perf(providers)!: make WatsonX SDKs opt-in promptfoo/promptfoo
WatsonX and IBM's core SDK no longer arrive by default; both move to optional peers while staying available for development. Anyone running IBM models has to install @ibm-cloud/watsonx-ai and ibm-cloud-sdk-core explicitly, and the provider reports actionable guidance when the SDKs are missing or unsupported.
feat(providers): add TypeSafe Jev provider promptfoo/promptfoo
typesafe:<model> wraps TypeSafe's Jev, a model that answers typed questions (Noul, Score, Choice) with probabilities rather than generated text. That lets Jev back the classifier assertion and grade llm-rubric, mapping the rubric to Noul, or to Score when levels are set.
fix(providers): honor per-call cache overrides promptfoo/promptfoo
AI21, Cohere, and LocalAI chat and completion providers now honor per-call cache overrides, with bustCache taking precedence over the legacy debug fallback, including an explicit bustCache: false. Default caching behavior is unchanged, so this only shifts what your existing flags already ask for.
fix(providers): launch npm-installed Codex app-server on Windows promptfoo/promptfoo
Windows cannot spawn npm's codex.cmd launcher directly, so the default command now starts npm-installed Codex through its JavaScript entrypoint and Node, with no shell and no new dependency. Global and project-local npm installs both work, and native Windows executable lookup is preserved.
fix(config): preserve imported test file path resolution promptfoo/promptfoo
Imported YAML tests, and JSON/JSONL tests loaded through arrays or globs, again resolve bare vars files and JavaScript/Python provider references from their source directory, restoring the behavior of release 0.123.1. Nested test-file references keep their own directory, so split configs that broke after that release should load again.
fix: preserve log context and validate media paths promptfoo/promptfoo
Structured Winston transports were dropping sanitized context fields; they now preserve those fields and caller location. Legacy local-media operations validate path components against the canonical storage root and use no-follow reads where the platform supports it, tightening what a crafted media path can reach.
MAINT: Bump mcp from 1.30.0 to 2.2.0 microsoft/PyRIT
PyRIT's MCP client moves from 1.30.0 to 2.2.0, a major-version jump on a dependency that many targets and orchestrators sit behind. Test your MCP-backed flows before pinning it in a production environment. Alongside: garak's Hugging Face InferenceAPI now raises a clear error on a 404 instead of a generic TypeError, and Codex app-server support now extends to Windows installs.
$ ls ai-red-teaming/week/ # the briefings behind this review
Keep up with AI Red Teaming in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
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.