Skip to main content
Opfor maintains two catalogs — agent and MCP. Derived suites are computed at load time from each evaluator’s standards: tags (see authoring); curated suites are hand-authored. The same evaluator can appear in several derived suites — it’s mapped once and shows up wherever its tags land.
Evaluators with a -source suffix (e.g. prompt-injection-source) are static source/sink code-analysis checks that pair with a dynamic sibling — they have no attack patterns and are skipped by the pattern-based judge pipeline.

Agent red-team

Suites

owasp-llm-top10

owasp-agentic-ai

owasp-mcp-top10 (agent-side)

Probes how an agent target behaves around MCP. To probe an MCP server directly, see the MCP catalog.

owasp-api-top10

eu-ai-act

nist-ai-rmf

mitre-atlas

output-trust-and-safety

Output-quality and trust-boundary subset (hallucination, sycophancy, impersonation, contractual overreach, off-topic drift, reasoning DoS, unicode-smuggled injection) that doesn’t map cleanly to one OWASP framework but is a real production liability. Curated.

harmful-content

Content-harm coverage from the MLCommons hazard taxonomy and Harmbench, plus targeted additions for hate, harassment, sexual content, fraud, disinformation, privacy, and copyright. Best paired with raw LLM endpoints; compose with owasp-llm-top10, owasp-api-top10, and eu-ai-act for full coverage. Curated.

pre-deploy-critical

Broad pre-deployment gate spanning the highest-severity failure modes — injection, leakage, unauthorized actions/access, and top harm categories. Curated; compose with the derived OWASP suites for full-standard coverage.

quick-smoke

Fast, high-signal subset for CI / a first run — one representative critical check across the main surfaces. Curated, intentionally small.

MCP red-team

Suites

Evaluators (23 pickable)

resource-exposure also runs automatically during opfor run Phase 0 — opfor calls resources/list + resources/read on every resource and judges for secret/PII exposure, independent of whether it’s in your selected suite/evaluator list. Disable with mcp.scanResources: false in the config. It’s a normal catalog member otherwise — pick it explicitly to get it into a run’s reported results outside the Phase 0 pre-flight. -source suffixed evaluators (path-traversal-source, secret-exposure-source, command-injection-source, ssrf-source, missing-authentication-source) are static source/sink code-analysis checks — no attack patterns, skill-driven, paired with their dynamic sibling.