Agentic AI

Browse prompts in this category.

Prompts

Agentic AI Security Architect
Design defense-in-depth for autonomous AI agents covering the OWASP Top 10 for Agentic Applications (2026) — trust boundaries, tool sandboxing, credential isolation, memory provenance, inter-agent mTLS, kill switches. Use when architecting agent platforms that plan, use tools, and act.
ASI01 Agent Goal Hijacking Defense
Harden AI agents against OWASP Agentic ASI01 Goal Hijacking via prompt injection in RAG, emails, and web content — input trust tiers, injection classifiers, goal anchoring, retrieved-content sandboxing, drift detection. Use when defending agents that process untrusted documents or external data.
ASI02 Tool Misuse and Exploitation Defense
Harden agent tool use against OWASP Agentic ASI02 — least-privilege policies, JSON Schema argument validation, chaining controls, description integrity pinning, sandboxed execution, human approval for destructive calls. Use when exposing tools to agents or reviewing tool chains.
ASI03 Identity and Privilege Abuse Defense
Harden agent identity against OWASP Agentic ASI03 — unique per-agent identity, JIT task-scoped tokens, ABAC least-privilege, confused-deputy prevention, session binding, non-transitive delegation. Use when designing agent IAM or auditing credential handling across agent chains.
ASI04 Agentic Supply Chain Security
Harden agentic supply chains against OWASP Agentic ASI04 — Agent-BOM inventory, signed manifests, MCP registry verification, tool/plugin sandboxing, prompt-template integrity, model-weight provenance, kill switches. Use when vetting agent components, plugins, or MCP servers for production.
ASI05 Unexpected Code Execution Defense
Harden agent code execution against OWASP Agentic ASI05 — gVisor/Firecracker/Wasm sandboxes, command-injection prevention, parameterized queries, safe deserialization, template isolation, and pre-exec static analysis. Use when agents generate or execute shell, SQL, or scripting code.
ASI06 Memory and Context Poisoning Defense
Harden agent memory and RAG stores against OWASP Agentic ASI06 poisoning — tenant namespace isolation, ingestion injection filtering, provenance tracking, embedding drift detection, summary integrity checks. Use when persisting agent memory, embeddings, or long-term context.
ASI07 Insecure Inter-Agent Communication Defense
Harden inter-agent messaging against OWASP Agentic ASI07 — mTLS, per-message signatures, anti-replay nonces/timestamps, delegation scope validation, schema enforcement, authenticated discovery registries. Use when designing multi-agent coordination, A2A protocols, or delegation flows.
ASI08 Cascading Failure Defense
Implement cascading-failure containment per OWASP Agentic ASI08 — circuit breakers, rate limits, plan checkpointing/rollback, stage-boundary output validation, feedback-loop detection, chaos-engineering. Use when orchestrating multi-step plans or multi-agent workflows with blast-radius concerns.
ASI09 Human-Agent Trust Exploitation Defense
Harden human-agent interfaces against OWASP Agentic ASI09 trust exploitation — AI-generated labeling, confidence indicators, critical-action confirmation with cooling-off, anti-urgency/authority controls, generated-code diff review. Use when designing agent UIs, approval flows, or code review.
ASI10 Rogue Agent Defense
Implement rogue-agent detection and containment per OWASP Agentic ASI10 — behavioral baselines, hash-chained audit trails, scope-drift detection, kill switches, self-replication prevention, DLP egress monitoring, collusion graphs. Use when running long-lived autonomous agents in production.
MCP Authentication Authorization and Session Security
Implement MCP authorization per spec — OAuth 2.1 with PKCE, audience validation, no token passthrough, per-client consent, exact redirect URI matching, CSPRNG user-bound sessions, progressive scope elevation, frame-ancestors consent hardening. Use when building MCP auth or reviewing OAuth flows.
MCP Input Validation and Prompt Injection Defense
Harden MCP server inputs — JSON Schema with additionalProperties false, prompt-injection pattern/classifier detection, allowlisted parameterized commands, canonical path resolution, parameterized SQL/NoSQL, output-context encoding. Use when MCP tools reach databases, shells, or filesystems.
MCP Network Security and SSRF Prevention
Harden MCP servers against SSRF and exfiltration — RFC 9728 private IP blocking on OAuth discovery, DNS pinning against rebinding, egress proxies like Smokescreen, TLS 1.3 with cert pinning, per-hop redirect validation, outbound DLP. Use when MCP servers fetch URLs or make outbound HTTP.
MCP Server Secure Development Architect
Design secure-by-default Model Context Protocol servers — stdio/mTLS transports, OAuth 2.1 with audience validation, CSPRNG user-bound sessions, strict JSON Schema tool validation, SQL/path/command injection prevention, and sanitized error handling. Use when building or hardening MCP servers.
MCP Supply Chain and Registry Security
Harden MCP ecosystems — curated server registries with cert pinning, signed tool manifests with quarantine, hash-pinned deps with CVE alerting, reproducible CI/CD with cosign, local-server consent dialogs, file-integrity monitoring. Use when managing MCP fleets or vetting community MCP servers.
MCP Tool Poisoning and Description Integrity Defense
Harden MCP tool metadata against poisoning — signed manifests with hash verification, NLP scans for imperative/role-switching language, strict schemas with additionalProperties false, quarantined registration, rug-pull drift detection. Use when operating MCP hubs or auditing third-party tools.