investigating how ai safety properties like alignment, robustness, interpretability degrade in low-resource and tonal language contexts.

bridging a decade of research, this review analyzes 70+ papers to identify the nine failure modes currently threatening AI safety.

addressing the challenge of LLM hallucinations in RAG systems, this paper explores simple, interpretable uncertainty signals for real-time detection.

by embedding strict safety standards into public tenders, african governments can instantly transform ai safety from a passive policy goal into an enforceable market requirement.

a vignette demonstrating how a secretly loyal ai assistant can cause state lock-in while evading behavioral audits, highlighting the effectiveness of weight-space provenance checks over naive per-decision audits.

comparing model weights easily detects hidden loyalties that behavioral tests completely miss, suggesting a necessary shift in AI auditing priorities.

an open-source protein-aware screening tool that catches adversarial DNA sequences evading standard hamming-distance checks by translating to amino acid homology, defending against context-scrubbed multi-agent LLM biosynthesis attacks.

research compares LSTM neural networks against Arps hyperbolic decline models for oil production forecasting using the volve field dataset.