$ the-wire · showcase
PyRIT adds tool role-play attacks, garak moves probe intents into reusable stubs
By RepoJournal · Filed · About AI Red Teaming · Composed from the cited sources · methodology
PyRIT shipped a new tool role-play converter and technique for red-teamers going after agentic tool use, while garak extracted the prompts baked into three baseline probes into reusable stubs without breaking existing detector mappings.
FEAT: Add tool role-play converter and technique support microsoft/PyRIT
PyRIT now ships a tool role-play converter and matching technique, so attacks can be constructed around simulated tool invocation rather than plain conversation turns. It is the first new offensive surface added to the converter set in this batch, and it is the change most worth trialling against agents that call functions.
feat(cas): extract probe intents into reusable stubs NVIDIA/garak
garak pulled the prompt content out of three no-technique probes into reusable stubs: ansiescape.AnsiRaw now points at S008terminal, apikey.GetKey at S009cred, and malwaregen.SubFunctions at S008code. The legacy probes and their detector mappings are preserved, so existing scan configs keep resolving as before; the stubs live in garak/intents/S008.py and garak/data/cas/intent_stubs/S009cred_get...
FIX: keep the conversation draft clean when applied conversions match no piece microsoft/PyRIT
A conversion map whose pieces are no longer in the draft used to make applyConversions rebuild every message and piece anyway, leaving the conversation permanently dirty behind the discard confirmation. It now tracks a changed flag and returns the original arrays when nothing matched, so discarding a proposal actually clears it.
feat: Add bounded Scenario result hierarchy projections (#3145) microsoft/PyRIT
Bounded scenario result hierarchy projections land in PyRIT, giving scenario runs a projected view of their result tree rather than the full hierarchy. Check the PR's breaking-change note before pinning it into an existing reporting pipeline.
FIX Improve single-turn chat controls and first-prompt branching microsoft/PyRIT
Single-turn chat controls and first-prompt branching get fixes, and the single-file initializer registry now uses descriptor-safe member inspection so unrelated dynamic attributes no longer get evaluated during discovery, while explicitly declared lazy exports still resolve; 62 registry tests pass.
