AI Governance Thought Leadership
AI Governance Thought Leadership: Safety, Compliance and Enterprise Risk Research
200+ papers. 50,000+ citations. Published at ICLR, ACL and EMNLP.
Key Findings on AI Compliance Risks
This is WalledAI's thought leadership hub for AI governance: peer-reviewed papers and analysis from our research team on LLM safety, red teaming, trustworthiness, compliance and enterprise AI risk. Every item below is published work you can read in full at its original venue. Three themes recur across the studies.
Safety Alignment Degrades After Fine-Tuning
Models that pass safety evaluation at release can lose those guarantees once they are adapted on domain data, so governance has to be re-tested after every model change.
Guardrails Break Under Adversarial Prompting
Red teaming work shows single-layer filters are bypassable. Controls need to sit at input, retrieval and output, not only on the prompt.
Safety Evaluation Is Language and Culture Dependent
Benchmarks built in one language under-report risk elsewhere, which matters for regulators that expect evidence for every market you operate in.
TL;DR
WalledAI's published research on making AI systems safer, more compliant and better prepared for real-world enterprise risk.
Showing 9 of 9 selected publications
Archive of AI Safety Studies
OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost Always!
A benchmark exposing how frequently LLMs respond to off-topic prompts in deployed assistant settings, and what that means for safety and reliability.
Read PaperMeasuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse
Introduces Trust-Score and Trust-Align to evaluate and improve LLM trustworthiness in RAG, with strong gains across ASQA, QAMPARI and ELI5.
Read PaperEvaluating AI for Finance: Is AI Credible at Assessing Investment Risk Appetite?
Assesses whether frontier LLMs can credibly evaluate investment risk appetite. Finds significant variance across GPT, Claude, Gemini, LLaMA and DeepSeek when user attributes change - violating AI and finance regulations.
Read PaperFerret: Faster and Effective Automated Red Teaming with Reward-Based Scoring Technique
A novel reward-based automated red teaming approach that generates adversarial prompts faster and more effectively than existing methods.
Read PaperRuby Teaming: Improving Quality Diversity Search with Memory for Automated Red Teaming
Enhances Rainbow Teaming with a memory cache dimension, achieving a 74% attack success rate - 20% higher than baseline - with improved quality diversity.
Read PaperLanguage Models are Homer Simpson! Safety Re-Alignment of Fine-tuned Language Models through Task Arithmetic
Proposes RESTA - a simple arithmetic method that restores safety in fine-tuned LLMs by adding a safety vector to compromised model weights, reducing harmfulness from 18.6% to 5.1% while preserving task performance.
Read PaperRED-EVAL: A New Paradigm for Red Teaming
Introduces a chain-of-utterances methodology for systematically red-teaming LLMs to discover safety vulnerabilities.
Read PaperLanguage Model Unalignment: Parametric Red-Teaming to Expose Hidden Harms and Biases
Exposes hidden biases and harmful outputs in aligned language models through parametric red-teaming techniques.
Read PaperInvestigating Gender Bias in BERT
Analyses gender bias in BERT's contextual embeddings across emotion and sentiment tasks. Proposes a fine-grained algorithm to identify and remove gender-specific directions from each BERT layer, significantly reducing downstream bias.
Read PaperWhat These Findings Mean for Enterprises
Model Approval Is Not a One-Time Event
Because alignment shifts with fine-tuning and vendor updates, an approved model is only approved for the version you tested. Continuous evaluation, not a launch-day sign-off, is what auditors under the EU AI Act and NIST AI RMF expect.
Controls Belong Outside the Model
Prompt-level filtering is the weakest link in every red teaming study. Enforcing policy in an independent runtime layer such as Walled Protect keeps the control in place when the underlying model is swapped.
Evidence Has to Be Reconstructable
Regulators ask what the system did on a specific date, for a specific user. That requires per-interaction logs of prompts, redactions and policy decisions, which is what the Governance Dashboard is built to produce.
Practical Guidance for AI Governance Teams
Start From an Inventory, Not a Policy
Map which teams use which AI tools and what data reaches them before writing rules. Classification work belongs first, which is why data classification precedes enforcement.
Redact Before the Prompt Leaves Your Perimeter
Masking identifiers at the edge with Walled Redact removes the largest single source of AI data-protection incidents.
Red Team on Your Own Data
Public benchmarks miss domain-specific failure modes. Run adversarial tests against your own prompts, retrieval corpus and languages of operation.
Full Publication Record of Our Founding Researchers
Together, our founders have published over 200 papers with 50,000+ citations. Browse their complete research on Google Scholar.
Applied Analysis on Enterprise AI Deployment
Practical insights, technical deep-dives and perspectives on AI safety, governance and enterprise deployment.
Visit the Blog