ForgeryGuard
Large language models will readily draft letters, certificates and official notices on request. That is useful, but it also makes forging institutional documents easier than it has ever been.
ForgeryGuard sits in front of the model and screens each request before it is answered. The aim is to block prompts that impersonate institutions or ask for fraudulent artifacts, while still allowing the legitimate document drafting that schools, offices and businesses depend on every day. The approach combines neural classification with symbolic reasoning, so that each decision can be explained rather than simply asserted.
Status: Research complete. Publication details will be added here once the paper is published.
Darry Derrick is a Zambian researcher working on the security of AI systems and on putting trustworthy AI to work in agriculture and education across Southern Africa. His work starts from a practical question: how do we make intelligent systems safe and dependable when they run on limited compute, patchy connectivity and mobile money rails, rather than inside a well resourced data centre?
He leads Darry Lab for Secure and Applied AI, an independent research initiative. Current work includes ForgeryGuard, a guardrail that stops large language models from producing forged institutional documents; QSentry, an open source tool for securing the quantum software supply chain; LUDA, a ledger verified crop diagnostics platform for smallholder farmers; and YieldTrace, a causal approach to attributing crop losses at farm level.