Open Source Summit India 2026

As banks shift from traditional software to LLMs, the threat landscape is evolving from "bugs in code" to "poison in data." Traditional vulnerability management (CVEs) cannot detect a model that has been trained to have a backdoor. This lightning talk explores the critical risk of Data Poisoning and Indirect Prompt Injection in a regulated fintech environment.

We will walk through concrete examples—from "hidden" instructions in customer documents to "Trojan Horse" models downloaded from public repositories—that can lead to unauthorized transfers or massive reputational damage. The session provides a 3-step governance framework for OSPOs to move beyond SCA and toward Model Integrity:

Implementing Data Lineage for fine-tuning sets,

Adopting Adversarial Red-Teaming as a standard release gate, a
Leveraging open-source frameworks like MITRE ATLAS to map AI-specific threats and tools like garak, augustus to detect the vulnerabilities
Learn why the OSPO is the natural home for AI Safety and how to protect your organization's "Intelligence Supply Chain" from being poisoned at the source.