RepoJournal
AI Red Teaming AI Red Teaming
29 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 2026-10-27
stories 39
RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

PyRIT validates penalty parameters, corrects GCG gradient aggregation

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

renovate varunj-msft nithin42 RohithPariki devangpratap romanlutz
6 people shipped this
renovate @renovate 1 cited varunj-msft @varunj-msft 1 cited nithin42 @nithin42 1 cited RohithPariki @RohithPariki 1 cited devangpratap @devangpratap 1 cited romanlutz @romanlutz

PyRIT landed a cluster of fixes that change how prompt targets reject bad parameter values and how GCG workers' gradients combine, while promptfoo's day was entirely dependency bumps.

FIX: Validate presence_penalty and frequency_penalty in OpenAI and LiteLLM prompt targets microsoft/PyRIT

by nithin42

OpenAIChatTarget, OpenAICompletionTarget, and LiteLLMChatTarget documented presence_penalty and frequency_penalty as accepting values from -2.0 to 2.0, but neither class validated them in __init__: out-of-range, NaN, or inf values were stored in target identifiers and only blew up downstream during the API request. The new validate_presence_penalty and validate_frequency_penalty helpers in pyri...

FIX GCG: only sum gradients of workers that share a vocabulary microsoft/PyRIT

by devangpratap

GCG now sums gradients only across workers that share a vocabulary, rather than aggregating across all of them. If you run GCG attacks split across models with differing tokenizers, the combined gradient was previously built from workers that could not meaningfully contribute to it.

FIX: Keep scorer parameters that name a server file out of the API microsoft/PyRIT

by varunj-msft

Scorer parameters that name a server-side file no longer leave the client and reach the API. Operators using server file references in scorer config should confirm the parameters they expect are still transmitted for the scorers they rely on.

FEAT GUI: Preserve chat drafts when exploring datasets (#2750) microsoft/PyRIT

by RohithPariki

The PyRIT GUI now preserves your in-progress chat draft while you browse datasets, so switching datasets mid-conversation no longer costs you typed input.

fix(deps): update dependency js-yaml to v5.5.0 promptfoo/promptfoo

by renovate

promptfoo's remaining work was housekeeping: js-yaml moved from 5.4.3 to 5.5.0, vitest from 5.0.4 to 5.0.5, and webpack from 5.111.1 to 5.112.0, with nothing to change in how you run the tool.

$ ls ai-red-teaming/reviews/ # the step-back reads

the week in review promptfoo 0.125.0 makes Claude and Codex SDKs opt-in
Full archive → Weekly reviews →

Quick answers

What shipped in AI Red Teaming on October 27, 2026?
PyRIT landed a cluster of fixes that change how prompt targets reject bad parameter values and how GCG workers' gradients combine, while promptfoo's day was entirely dependency bumps. In total, 17 commits and 22 pull requests landed.
Who contributed to AI Red Teaming on October 27, 2026?
6 developers shipped this update, including renovate, varunj-msft, romanlutz, nithin42, RohithPariki, and devangpratap.
What were the notable AI Red Teaming updates?
FIX: Validate presence_penalty and frequency_penalty in OpenAI and LiteLLM prompt targets, FIX GCG: only sum gradients of workers that share a vocabulary, and FIX: Keep scorer parameters that name a server file out of the API.
♥

$ log AI Red Teaming --recent

PyRIT tidies its warning UI and pins down converter state across prompt restores 26 shipped Anthropic SDK tracing collides with promptfoo spans, and PyRIT adds a new attack technique 37 shipped PyRIT adds tool role-play attacks, garak moves probe intents into reusable stubs 24 shipped Promptfoo moves off vulnerable MSAL, PyRIT adds a char-split attack and cloud token detection 69 shipped

$ log --all # elsewhere on the wire

Alignment & Post-training XPU rollout images, a tensordict 0.11 data fix, and a gated-judge guard Agent Security Denied RAG queries get real hashes, Docker deployer logs gain stderr

$ ls showcases/

Safety Evals Agent Security Interpretability Alignment & Post-training

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.