Cross-lingual guardrail auditing & inference steering tools.

Raising $15,000active

Ready-to-use SAE steering tools and standalone evaluation suites to patch cross-lingual jailbreak vectors in frontier deployment stacks.

Project Details

In our recent paper accepted at ICML 2026 GlobalSouthML workshop, we show that frontier models are broken across languages. They refuse harmful prompts in English but comply when those same prompts are inputted in low-resource languages like Yoruba, Igala, Hausa, Igbo, or Swahili. We then used a training-free mechanistic method called LSR-Anchoring that uses Sparse Autoencoders-derived activation steering to clamp the English refusal direction to the low resource language prompt at inference time. The method recovered safety successfully in Yoruba, Igbo, Igala, and Hausa without degradation in capabilities. Here's the link: My ICLM paper; https://icml.cc/virtual/2026/78353.

The paper published for phase 1 of the research: https://arxiv.org/abs/2603.19273

This project builds the production engineering tools to fix that problem. We're creating a unified suite of open-source tools to audit and patch these latent cross-lingual jailbreak vectors at inference time.

The focus is mostly on shipping functional engineering codes. We don't need expensive retraining, as labelled safety dataset is not even available at scale for these low-resource languages, hence, this training-free method. Instead, our tools inject language agnostic Sparse Autoencoders (SAE) features directly into model inference pipelines. This stops harmful refusal failure without significant degradation in benign capabilities. I am the solo engineer on this project, pairing with native speaker annotators to validate the evaluation suites.

What we want to deliver:

A cross lingual steering tensor library; pre-computed, optimized SAE feature IDs and dense steering weights for Llama-3.1-8b, Llama3.3-70b, Mistral-7b, and qwen-2.5-7b covering Arabic, Yoruba, Igbo, Igala, Hausa, and Swahili.
Automated guardrails evaluation benchmark plugged into the UK AISI inspect platform so external red-teamers can automatically stress-test models for cross-lingual safety collapse.
A technical paper containing deployment rules, interference thresholds, and calculated failure zones for safety engineers.

Theory of Impact

As AI becomes more powerful and racing towards AGI, English based guardrails do not matter much if a model has dangerous capabilities. If a bad actor can extract bio-weapon protocols or cyber-warfare information using any of these low-resource language tokens, then the model is unsafe. We can not solely rely on English guardrails as that leaves a catastrophic bypass wide open for bad actors to utilize.

This project provides defensive tools to help reduce this x-risk by putting verified steering tensors into automated evaluation frameworks like the UK AISI inspect platform. It will help researchers working on cross-lingual LLM safety and also help labs audit their system before deployment. As verified in our research paper published at ICML 2026, these SAE features help shut down cross-lingual jailbreak vectors directly at the weight layer during inference.

People

Godwin Abuh Faruna

Team Member

Grants Received

Add grant

ICML 2026 Travel Award

from BlueDot Impact

$3,500

2026 BlueDot Technical Project Grant

from BlueDot Impact

$800

Google TPU Research Cloud Credit Grant

from Google TRC (TPU Research Cloud)

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$15,000

Ideal

$30,000

How the money will be spent

The money goes to compute infrastructure, dataset construction, and technical publication costs. My personal runway is being handled by a separate project grant from Manifund.

The money goes into GPU rentals to run large scale activation sweeps, extracting SAE feature dictionaries across multiple model architectures up to 70b parameters, red-teaming data & native contractors, and tooling release & hosting costs for public model weights.

Compute ($14,000-$18,000). This will be spent on GPU rentals (runpod) to run large scale activation sweeps for extracting SAE dictionary paths for all the architectures mentioned above up to 70b parameters.
Red-teaming data & native contractors ($4,000-$6,000) to be used as stipend for native Arabic Yoruba, Igbo, Hausa, Igala, and Swahili speakers to manually audit prompts, checking for translation artifacts and validating the evaluation suite.
Hosting costs for open-source tensor libraries ($2,000 - $6,000).

Discussion

No comments yet. Be the first to share your thoughts.