How an AI agent is attacked and defended, in thirteen slides: the LLM as an attack surface, direct and indirect prompt injection, the lethal trifecta, data leakage, excessive agency, defences in layers, least privilege with a Kubernetes ServiceAccount, human approval for writes, input and output guardrails, the OWASP Top 10 for LLM Applications, governance records, and red-teaming in CI.
The slides
The tool is lesson 9 of
LLMOps for AI Agents, in thirteen slides:
- How do we secure an AI agent?: one instruction hidden in a log line, read by an agent with no limits and by a limited one.
- The LLM as an attack surface: one context window, where instructions and data are the same tokens.
- Direct and indirect prompt injection: a log line that tells the agent to open 50 tickets.
- The lethal trifecta: private data, untrusted content and a way out, in one agent.
- Data leakage: secrets, personal data, documents across teams, the system prompt.
- Excessive agency: the tools an agent holds: the on-call assistant's tools, and two it was never given.
- Defences in layers: seven defences on the three paths through the agent.
- Least privilege: the agent's own credentials: a read-only Role, and
kubectl auth can-i. - Human approval for writes: the run paused on
open_ticket, and the card a person sees. - Input and output guardrails: where the checks sit, and an answer redacted.
- OWASP Top 10 for LLM Applications: the 2025 list, read against the on-call assistant.
- Governance: records of models, data and actions: an audit-log entry and four other records.
- Red-teaming and security evals in CI: the pipeline, and a promptfoo red-team configuration.
The buttons on the left and right edges move between slides. So do the arrow keys and a swipe.
The lab
The log lines, the ticket, the audit-log entry and the redacted answer are made up for the drawings; the
placeholders in the redacted answer are an example, and each guardrail product writes its own.
Sources