1.6 KiB
1.6 KiB
AI Agent, Prompt Injection, Cloud, And Supply Chain
Use this reference for prompt-injection challenges, toolchain or MCP tasks, agent memory issues, containerized deployments, CI/CD, cloud control plane analysis, or supply-chain tracing.
Agent And Prompt Injection
- Treat prompts, tool schemas, retrieved chunks, planner notes, memory files, and handoff messages as challenge artifacts.
- Map the control stack: instruction layers, retrieval layers, memory layers, tool gates, auth material, and side effects.
- Prove one minimal exploit chain from untrusted content to model-visible instruction to tool side effect or exposed secret.
- Distinguish claimed capability from observed capability.
Cloud, Containers, And CI/CD
- Split build-time, deploy-time, and runtime.
- Reconcile checked-in manifests with live containers, mounted secrets, env, logs, and traffic.
- Trace provenance end-to-end: source -> dependency resolution -> build -> packaging/signing -> publish -> runtime consumer.
- Treat registries, metadata services, message buses, object stores, and IAM-looking identities as sandbox control surfaces when they appear in the challenge path.
Evidence To Keep
- Exact prompts, retrieved chunks, planner/tool transitions, and final tool args
- Compose/Kubernetes/Terraform snippets tied to live mounts or routes
- Artifact hashes, version drift, registry pulls, CI steps, and resulting runtime hooks
Common Pitfalls
- Treating a prompt string as authoritative without runtime confirmation
- Confusing infrastructure intent with live deployment truth
- Missing the boundary where a retrieved chunk becomes executable tool input