CLI Reference

CLI Reference

The console script is aphasia. Run aphasia --help for the live version.

aphasia [--version] <command> ...

Commands: list, payloads, run, report.

aphasia list

Prints the scenario catalogue (OWASP LLM Top 10) as a table: Scenario, OWASP, ATLAS, Channel, Seeds, Category. See Scenarios.

aphasia payloads

Prints how many seed injection payloads ship per scenario. See Seed Payload Library.

aphasia run

Runs scenarios against a target, writes a report + evidence log, and returns an exit code.

Flag	Default	Meaning
--target	fixture	fixture (bundled agent on 127.0.0.1:8000), an http(s):// chat URL, or an MCP endpoint. See Targets.
--scenario	all	Run only this scenario id. Repeatable.
--mode	hybrid	seed / adaptive / hybrid. See Attacker Modes.
--model	ollama/llama3.2:1b	litellm model string for the attacker. See LLM Providers.
--api-base	$OLLAMA_API_BASE or $OLLAMA_HOST	Base URL for a self-hosted backend (Ollama/vLLM/LM Studio). Unused in seed mode.
--max-steps	6	Max attacker steps per scenario.
--out	runs/<timestamp>	Output dir for report.html, report.json, evidence.jsonl.

Exit codes

• 0 — no scenario proven.
• 1 — at least one scenario proven (success or metered). Use this to gate CI.
• 2 — a target/provider error occurred (and nothing was proven).

aphasia report --run-dir <dir>

Re-renders report.html / report.json from a past run's evidence.jsonl. Note: final capped / mirror_only verdicts are not stored as rows, so a re-render shows per-step verdicts; the live run report is faithful. See The Report.

Examples

aphasia run --mode seed                                  # no LLM, vs the fixture
aphasia run --model gpt-4o-mini                          # adaptive, OpenAI
aphasia run --scenario rag_exfil --scenario data_exfil   # a subset
aphasia run --target http://host:8000/chat --max-steps 10
aphasia report --run-dir runs/20261005T101533-74aa0d