Loading ToolBench
LLM security red teaming plugin + MCP server for testing prompt injection, jailbreaks, tool abuse, PII leakage, bias, and more
PromptArmor has 9 well-named tools with clear action verbs (scan, list, get, delete, generate, run, configure, check). All tools have descriptions (avg ~120 chars) and Zod schemas with typed parameters. However, several parameter descriptions are generic or missing context. Output schemas are not documented, LLMs cannot predict response structure. Error handling returns raw error messages without recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Parameter descriptions lack format constraints and dependency hints.
Check the health and status of the PromptArmor server.
Validate and preview the configuration for a target endpoint. Shows how requests will be formatted.
Delete a security scan record by its ID.
Generate a security report for a completed scan in text, JSON, or SARIF format.
Retrieve detailed results for a specific security scan by its ID.
List all available attack plugins with their metadata, categories, and severity levels.
List all previous security scan records, newest first.
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination). Scan results, plugin metadata, and report formats are opaque.
No tool annotations (readOnlyHint, destructiveHint). prompt_armor_delete_scan is destructive but unmarked; prompt_armor_scan makes external HTTP calls but not annotated. Agents cannot reason about side effects.
Error handling returns raw error messages without recovery guidance. 'Scan failed: <error>' tells LLM nothing about retry eligibility, user action, or next steps.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Run a specific attack plugin against a target endpoint. Useful for testing individual vulnerability types.
Run a comprehensive LLM security scan against a target endpoint. Tests for prompt injection, jailbreaks, tool abuse, PII leakage, bias, and more.
Parameter descriptions lack format constraints and dependency hints. 'url' accepts any string; 'format' enum is present but no guidance on when to use each. 'requestField' and 'responseField' lack JSON path examples.
No pagination guidance. prompt_armor_list_scans and prompt_armor_list_plugins accept 'limit' but no 'offset', 'cursor', or 'total_count' documented. Large result sets risk context window exhaustion.