$ cat ai-red-teaming/week/2026-11-16.log
the week in review · Nov 16 – Nov 22, 2026
promptfoo 0.126.0 restricts derivedMetrics formulas
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
PyRIT also marks two changes breaking, and code-scan-action 0.2.3 pins the filesystem MCP server to the GHSA-patched 2.3.0.
0.126.0 promptfoo/promptfoo
by promptfoobot[bot]
The release's own before-upgrading note is the thing to read: arithmetic, comparisons, ternaries and score names like __count still evaluate, but function calls and property access in derivedMetrics strings are out. Audit those formulas in your configs before you pull the tag.
fix(code-scan): upgrade the filesystem MCP server to 2.3.0 promptfoo/promptfoo
The optional filesystem MCP server is pinned to 2.3.0, the first release patched for GHSA-p9cv-8g75-4xcg, and since it declares Zod, strict pnpm and Yarn installs can drop the old package extension. Yarn Plug'n'Play still needs its ESM loader. Full-repository scans are the affected path, so operators running them should take the bump.
code-scan-action: 0.2.3 promptfoo/promptfoo
by promptfoobot[bot]
This is the release that carries the MCP pin for full-repository scans, and eligible committed text files from the compared revision can now be shared with the hosted scanner. Operators running repo-wide scans get the fix without patching manifests by hand.
fix(redteam): isolate live config files for concurrent runs promptfoo/promptfoo
Two live red-team runs started in the same millisecond could write the same YAML before generation read it, letting one run evaluate the other's config. Each live config now gets a UUID filename written exclusively with 0600 permissions on POSIX. If you run concurrent red-team jobs, this was silently corrupting results.
[BREAKING] FIX: Do not score write-only target deliveries as attack responses (#3757) microsoft/PyRIT
Write-only target deliveries are no longer scored as attack responses, which changes results for any target that acknowledges without returning content. The PR is marked breaking, so your historical scores against such targets may not be comparable with runs after the upgrade.
[BREAKING] FIX: Resolve Pydantic field collisions and deprecated dependency interfaces microsoft/PyRIT
Pydantic field collisions and deprecated dependency interfaces are handled here, and the PR is flagged breaking. Set aside a test pass on your target and converter configs rather than assuming a clean drop-in.
FEAT: Fetch trace evidence from a remote OTLP/JSON backend microsoft/PyRIT
OtlpJsonTraceClient joins the trace layer as a TraceClient that fetches evidence from a remote OTLP/JSON backend, the remote adapter Phase 5 had deferred. Useful if you already export traces to a collector and want scoring to read from there instead of local memory.
feat(providers): support GPT-6.2 Astra and Sol promptfoo/promptfoo
Operators can now point configs at openai:gpt-6.2-astra or openai:gpt-6.2-sol for Responses, including tools, with the explicit openai:chat: prefix selecting Chat Completions for text. Alongside: garak opened an opt-in agent_breaker.SourceToSink probe for bounded multi-tool chain exploration, and PyRIT added an opt-in prefill technique with native assistant continuation across Hugging Face, Ope...
$ 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.