AI, AppSec, and You: A Practitioner's Diary

The session asks whether AI will reshape application security the way cloud did, or whether it is an overheated bubble. The speaker, a DefectDojo co-founder, answers from his own experiments rather than predictions. He also raises the economics behind large AI vendors, including token costs, heavy data center spending and the risk that a few providers dominate the market.

The central case study is an MCP server built to expose DefectDojo findings to language models. The main lesson was that input quality matters more than the connector. Because DefectDojo already normalizes results from many scanners, the model receives one consistent view of a finding. Other practical points: send only the fields the model needs, since every token costs money; use result limits so a count query returns a single number; keep the server read-only; and treat tool descriptions as instructions the model will follow. The talk also covers running open-source models locally with Ollama and orchestration libraries such as LangChain, LangGraph and LangFlow. It explains how errors compound across multi-step agent chains, and how a DARPA entry placed deterministic checks between steps to keep accuracy up. Prompt injection, non-deterministic outputs and the lethal trifecta of untrusted input, sensitive data access and outbound communication are presented as design hazards.

The takeaways are to avoid tying a company to one AI vendor, to add guardrails and safety margins, to limit blast radius and to fail safely. The speaker closes with open problems he would like solved, such as linking findings to developer workflows and handling changing vendor output formats automatically.