$ cat ai-red-teaming/week/2026-11-09.log
the week in review · Nov 9 – Nov 15, 2026
PyRIT v1.2.0 overhauls CoPyRIT GUI for scenario reuse
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
The release lets operators save scenario configurations persistently, and promptfoo brings Claude Fable and Mythos 5.2 across four providers.
v1.2.0 microsoft/PyRIT
Reusable scenario configurations, grouped launches, persistent engagements and findings, and richer result analytics land in the GUI, so assessment runs no longer have to be re-set-up from scratch each time. The release is marked breaking, so read the changed surface before upgrading if you script or extend the GUI.
feat(providers): add Claude Fable and Mythos 5.2 promptfoo/promptfoo
Adds `claude-fable-5-2` and `claude-mythos-5-2` on the Anthropic API, Amazon Bedrock, Google Vertex AI and Azure AI Foundry, with most 5.1 references in docs, examples and UI pickers retargeted. A new `CL...` request rule is introduced, so check your Anthropic target params before pointing a red-team config at the new models.
FEAT: Add SecretDisclosureScorer for partial and obfuscated disclosure of a known secret microsoft/PyRIT
The new float-scale `SecretDisclosureScorer` measures how much of a known secret a response reveals, in bits, searching the response as written, with separators removed, and with digit words and NATO forms handled. That makes partial and obfuscated leaks of recovery codes, keys and account numbers gradeable instead of binary pass/fail.
FEAT: Bound OpenAIResponseTarget's tool loop, and end the turn when a tool's infrastructure fails (part 1) microsoft/PyRIT
`OpenAIResponseTarget`'s tool loop is now bounded, and a turn ends when a tool's infrastructure fails rather than continuing the loop. Part 1 of 5 of the sandboxed code execution series, and the release notes-style split means it works without a sandbox.
FIX: send converter instructions in the user turn so converters write nothing to memory microsoft/PyRIT
LLM-backed converters now send their instructions in the user turn, so converter requests and responses stay out of the target's conversation history. If you rely on converter traffic not polluting scored conversations, this changes what lands in memory.
FEAT: Add FullwidthConverter microsoft/PyRIT
`FullwidthConverter` maps printable ASCII to Unicode fullwidth forms (U+FF01 to U+FF5E), normalizing back to the original under NFKC. It is a probe for filters and classifiers that match exact characters rather than normalized text.
fix(providers): replay video cache hits from one verified mapping promptfoo/promptfoo
Cache hits now replay from the single mapping whose primary video was verified, so OpenAI builds the cached video, thumbnail and spritesheet from the same snapshot and Azure and xAI use its video key. Previously a mapping replaced between reads could return video A alongside B's assets.
[FIX] Assign local YAML filename fallback to seeds so named queries find them microsoft/PyRIT
Alongside: local YAML datasets that omit `dataset_name` now get the filename fallback on their seeds too, so `get_seeds_async` with the stem finds them; JSON conversation exports and HTML report pieces carry the converted data type; and a saved-draft fix keeps unsent chat input per conversation when you navigate or switch.
$ 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.