Modern Appsec vs. GenAI Application: Is Your Appsec Ready?

The talk starts from the rapid spread of generative AI features and AI code assistants in software teams, and asks whether a mature application security program is equipped for the new risks. The speaker sketches a generic GenAI application stack, from the hosting infrastructure and foundation model up through retrieval-augmented generation, plugins and agents, which gives the audience a shared map of where risk sits.

Next, the speaker walks the OWASP Top 10 for LLM applications against that stack and rates each item by how much existing controls already cover it. Prompt injection and insecure output handling mostly map to familiar input validation, privilege separation and web defenses, though with LLM-specific adjustments. Training and retrieval data poisoning, hallucination-driven over-reliance and excessive agency are presented as areas needing new controls, such as data provenance checks and human review of autonomous actions. Denial of service, supply chain and sensitive data disclosure are treated as extensions of current practice at larger scale.

The second half covers code produced by AI assistants, which can carry insecure patterns learned from old public code, outdated libraries, licensing problems, hallucinated or typosquatted packages, and misconfigurations. The recommended response combines AI usage governance, mandatory peer review, modern software composition analysis, developer training, monitoring of assistant use, and thorough testing of generated code. The practical takeaway is that most existing appsec controls still apply, while data handling, agent autonomy and generated-code governance need deliberate additions.