Securing AI/LLM Workloads in CI/CD: Automated Risk Mapping and Attack Simulation in Cloud Pipelines

Most teams integrating LLMs into cloud applications are flying blind in their pipelines. They’ve added scanners, maybe some guardrails, but they don’t actually know if those controls hold up against real attacks. This talk walks through a practical approach I’ve been working on to close that gap. The idea is simple: treat AI/LLM security like any other part of the pipeline—something you can measure, test, and break before it gets to production. I’ll show a working setup that ingests CI/CD configurations (GitHub Actions, GitLab CI) and builds a basic risk picture: what controls are missing, where data could leak, and how exposed the LLM layer actually is. It’s not perfect—mapping pipelines to meaningful risk is messy, and there’s a fair amount of noise—but it’s enough to surface the problems most teams aren’t seeing. From there, we layer in attack simulation. Instead of assuming prompt filtering or guardrails work, we actively try to break them during the pipeline. Prompt injection, data exfiltration, and system prompt leakage are all tested as part of the build. In a lot of cases, the first run fails—which is the point. Once controls are added or tightened, we re-run the same tests and use that as the signal for whether a build should pass. It turns security from a checklist into something closer to validation. This isn’t a silver bullet. LLM behavior is inconsistent, false positives happen, and scaling this across multiple pipelines takes effort. But it’s a practical way to move from “we think it’s secure” to “we’ve actually tested it.” Attendees will leave with a concrete pattern they can adapt: how to map pipeline risk, how to simulate real LLM attacks, and how to wire both into CI/CD without slowing teams down. Learning Objectives • Identify where AI/LLM risks show up in CI/CD pipelines. • Build a pipeline risk mapping approach to highlight weak points. • Add prompt injection and data exfiltration testing into workflows. • Use attack results as a gating signal. • Understand tradeoffs and limitations. Session Outline (30 Minutes) • Where pipelines break for AI security (5 min) • Mapping risk from pipeline configs (8 min) • Breaking the system on purpose (10 min) • Turning it into a validation loop (5 min) • What didn’t work (2 min)