The Shadow AI Self-Audit: Five Questions to Run on Your Own Organisation
A five-question self-audit you can run on your own company in about five minutes. No tool, no install, no gate. Find out where your AI audit trail actually stops.
Papers
What we found, written down
Reports and research from the team building Aisty: cost models, audit and assurance, isolation architecture, and what the regulation actually asks for. Working notes, with the numbers attached.
The AI Incident Timeline: Fifteen Failures, One Shape
A dated record of AI security incidents from 2023 to 2026. Prompt injection, data exfiltration, rogue agents, misconfiguration. The pattern is not a series of outliers. It is the baseline, and it is accelerating.
Long-Context Degradation
Effective context length under literal retrieval, semantic retrieval, multi-hop reasoning, conflict resolution, and grounded long-form generation.
Focus-Guided Transformers: Selective Context Attention for Efficient Code-Agent Reasoning
Large language models deployed as coding agents must process increasingly large codebase contexts, but standard autoregressive attention scales quadratically with sequence length. We introduce the Focus-Guided Transformer, which augments pretrained decoder-only models with a lightweight structural attention mechanism over code-context chunks, achieving 70–91% context reduction while maintaining or improving answer quality on 67% of coding-agent queries.
CODA: Concept-Operation Decoupled Architecture for Knowledge-Transparent Reasoning
CODA is a neural architecture that cleanly separates reasoning operations from factual knowledge. Knowledge lives in an external, structured memory that can be swapped at inference time without retraining. With memory the model reaches 63.7% accuracy against a 3.8% random baseline; without it, accuracy collapses to 1.4%. Swapping between two entirely different knowledge bases leaves performance identical at 87.5%, confirming that the weights encode only reasoning operations.