AI Agent Instructions — Prisma AIRS CLI

This document instructs AI agents (Claude Code, Gemini CLI, etc.) on how to use the airs CLI to interact with Palo Alto Prisma AIRS. It is a machine-readable reference — not user documentation. For human docs see docs/ or https://cdot65.github.io/prisma-airs-cli/.

Quick Context

airs is a CLI for Palo Alto Prisma AIRS AI security platform. It covers:

Runtime Security — scan prompts against AIRS security profiles, manage profiles/topics/API keys
Guardrail Optimization — atomic CLI commands for custom topic guardrails, driven by an external agent loop (see program.md)
AI Red Teaming — adversarial scans against targets with static/dynamic/custom attacks
Model Security — ML model supply chain scanning, security groups, rules, violations
Profile Audits — multi-topic evaluation with conflict detection
Backup & Restore — export/import AIRS configuration (targets, etc.) to/from local JSON/YAML files

The binary is airs. Five top-level command groups: runtime, redteam, model-security, backup, restore. Global flag --debug logs all AIRS/SCM API requests and responses to ~/.prisma-airs/debug-api-<timestamp>.jsonl.

Authentication

Different commands require different credentials. Set these as environment variables or in ~/.prisma-airs/config.json.

Credential Sets

Credential Set	Environment Variables	Used By
Scanner API	`PANW_AI_SEC_API_KEY`	`runtime scan`, `runtime bulk-scan`, `runtime topics eval`, `runtime profiles audit`
Management API (OAuth2)	`PANW_MGMT_CLIENT_ID`, `PANW_MGMT_CLIENT_SECRET`, `PANW_MGMT_TSG_ID`	All CRUD commands (profiles, topics, api-keys, customer-apps), all redteam commands, all model-security commands, `backup`, `restore`
LLM Provider	Depends on provider (see below)	`runtime profiles audit`

LLM Providers

Provider Value	Required Env Vars	Default Model
`claude-api`	`ANTHROPIC_API_KEY`	`claude-opus-4-6`
`claude-vertex`	`GOOGLE_CLOUD_PROJECT`	`claude-opus-4-6`
`claude-bedrock`	`AWS_REGION`, `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`	`anthropic.claude-opus-4-6-v1`
`gemini-api`	`GOOGLE_API_KEY`	`gemini-2.5-pro`
`gemini-vertex`	`GOOGLE_CLOUD_PROJECT`	`gemini-2.5-pro`
`gemini-bedrock`	`AWS_REGION`, `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`	`gemini-2.5-pro`

Optional Management Endpoints

Variable	Purpose
`PANW_MGMT_ENDPOINT`	Override management API base URL
`PANW_MGMT_TOKEN_ENDPOINT`	Override OAuth2 token URL

Verifying Credentials

# Scanner API — should return scan result
airs runtime scan --profile <profile-name> "test prompt"

# Management API — should return list of profiles
airs runtime profiles list

# Red Team — should return targets
airs redteam targets list

# Model Security — should return groups
airs model-security groups list

Output Formats

All list commands accept --output <format>:

Format	Description	Best For
`pretty`	Colored terminal output (default)	Human reading
`table`	ASCII table with box-drawing chars	Human reading, aligned columns
`csv`	RFC 4180 CSV with header row	Piping to files, spreadsheets
`json`	Pretty-printed JSON array	Programmatic parsing by agents
`yaml`	YAML with `---` separators	Config files, readability

For AI agents: always use --output json when you need to parse results programmatically.

# Get parseable JSON output
airs runtime profiles list --output json
airs redteam targets list --output json
airs model-security groups list --output json

Command Reference

Runtime — Scanning

Scan a single prompt

airs runtime scan --profile <profile-name> "<prompt>"
airs runtime scan --profile <profile-name> --response "<response>" "<prompt>"

Required: --profile, <prompt> argument Optional: --response (scan prompt+response pair) Auth: Scanner API key

Output fields: Action (block/allow), Category, Triggered (yes/no), Scan ID, Report ID, Detections list

Bulk scan from file

airs runtime bulk-scan --profile <profile-name> --input <file> [--output <csv>] [--session-id <id>]

Required: --profile, --input Input formats: .csv (extracts prompt column by header), .txt (one prompt per line) Output: CSV file with results (default: <profile>-bulk-scan.csv) Auth: Scanner API key

Behavior: Batches prompts in groups of 5, submits async, polls with exponential backoff on rate limit. Saves scan IDs to ~/.prisma-airs/bulk-scans/ for crash recovery.

Resume polling

airs runtime resume-poll <stateFile> [--output <csv>]

Resumes polling from saved bulk scan state file after a crash or rate limit failure.

Generate DLP test files

airs runtime dlp generate [--types pdf,png,jpeg,svg,docx] [--count <n>] [--out <dir>] [--techniques all|<ids>] [--seed <n>] [--output pretty|json]

Generates clean carrier files and "dirty" copies with synthetic sensitive data embedded via multiple techniques (per format). Writes <out>/clean/, <out>/dirty/, and <out>/manifest.json (dirty file → technique + embedded values, for scoring). No auth required (local file generation). Technique ids: PDF meta|hidden-text|trailer|visible|visible-samecolor; PNG text-chunks|trailer|stego-lsb|visible; JPEG exif|com|trailer|visible; SVG meta|hidden-text|comment|visible; DOCX core-props|hidden-run|visible|visible-samecolor. visible = foreground ≠ background; visible-samecolor (PDF & DOCX) = foreground == background (camouflaged but extractable). All values are synthetic / reserved-for-testing.

Runtime — Configuration Management

All CRUD commands require Management API credentials.

Security Profiles

airs runtime profiles list [--limit <n>] [--offset <n>] [--output <format>]
airs runtime profiles get <nameOrId> [--output <pretty|json|yaml>]
airs runtime profiles create --name <name> [protection flags...]
airs runtime profiles update <profileId> [protection flags...]
airs runtime profiles delete <profileId> [--force --updated-by <email>]

create requires --name. All protection flags are optional — omitted sections get AIRS defaults.

update uses read-modify-write: fetches the current profile, merges only the flags you specify, PUTs the full payload. Existing policy sections (including topic-guardrails) are preserved. Only specify what you want to change.

Profile protection flags (available on both create and update):

Flag	Description
`--prompt-injection <action>`	Prompt injection action (`block`/`allow`/`alert`)
`--toxic-content <action>`	Toxic content action (e.g. `"high:block, moderate:block"`)
`--contextual-grounding <action>`	Contextual grounding action
`--malicious-code <action>`	Malicious code protection action
`--url-action <action>`	URL detected action
`--allow-url-categories <list>`	Comma-separated URL categories to allow
`--block-url-categories <list>`	Comma-separated URL categories to block
`--alert-url-categories <list>`	Comma-separated URL categories to alert
`--agent-security <action>`	Agent security action
`--dlp-action <action>`	Data leak detection action
`--dlp-profiles <list>`	Comma-separated DLP profile names
`--mask-data-inline`	Mask detected data inline (boolean)
`--db-security-create <action>`	Database create action
`--db-security-read <action>`	Database read action
`--db-security-update <action>`	Database update action
`--db-security-delete <action>`	Database delete action
`--inline-timeout-action <action>`	Inline timeout action (`block`/`allow`)
`--max-inline-latency <n>`	Max inline latency in seconds
`--mask-data-in-storage`	Mask data in storage (boolean)
`--no-active`	Create/set profile as inactive

Examples:

# Create with multiple protections
airs runtime profiles create --name "Prod Firewall" \
  --prompt-injection block \
  --toxic-content "high:block, moderate:alert" \
  --malicious-code block \
  --agent-security block

# Update only one setting (everything else preserved)
airs runtime profiles update <profileId> --toxic-content "high:alert"

Custom Topics

airs runtime topics list [--limit <n>] [--offset <n>] [--output <format>]
airs runtime topics get <nameOrId> [--output pretty|json|yaml]
airs runtime topics create --name <name> --description <desc> --examples <ex1> <ex2> [--format json]
airs runtime topics apply --profile <name> --name <name> --intent <block|allow> [--format json]
airs runtime topics eval --profile <name> --prompts <csv> --topic <name> [--format json]
airs runtime topics revert --profile <name> --name <name> [--format json]
airs runtime topics sample [--output <path>]
airs runtime topics update <topicId> --config <json-file>
airs runtime topics delete <topicId> [--force --updated-by <email>]

API Keys

airs runtime api-keys list [--limit <n>] [--output <format>]
airs runtime api-keys create --config <json-file>
airs runtime api-keys regenerate <apiKeyId> --interval <n> --unit <unit> [--updated-by <email>]
airs runtime api-keys delete <apiKeyName> --updated-by <email>

--unit values: hours, days, months

Customer Apps

airs runtime customer-apps list [--limit <n>] [--output <format>]
airs runtime customer-apps get <appName>
airs runtime customer-apps update <appId> --config <json-file>
airs runtime customer-apps delete <appName> --updated-by <email>

Deployment Profiles (read-only)

airs runtime deployment-profiles list [--unactivated] [--output <format>]

DLP Profiles (read-only)

airs runtime dlp-profiles list [--output <format>]

Scan Logs

airs runtime scan-logs query --interval <n> --unit <unit> [--filter <all|benign|threat>] [--page <n>] [--page-size <n>] [--output <format>]

Required: --interval, --unit

Runtime — Guardrail Optimization

The guardrail workflow uses atomic commands designed for external agent loops (see program.md):

Create or update a topic

airs runtime topics create --name <name> --description <desc> --examples <ex1> <ex2> [--format json]

Validates AIRS constraints (name length, description length, example limits) and upserts by name. If a topic with the same name exists, it is updated.

Auth: Management API

Assign topic to profile

airs runtime topics apply --profile <name> --name <name> --intent <block|allow> [--format json]

Additive — preserves existing topics already assigned to the profile.

Auth: Management API

Evaluate topic against prompt set

airs runtime topics eval --profile <name> --prompts <csv> --topic <name> [--format json]

Scans a static CSV prompt set against the profile, computes metrics (TPR, TNR, coverage, F1), and returns FP/FN details.

CSV format: Three required columns: prompt, expected (belongs to topic: true/false), intent (block/allow). All rows must have the same intent. Run airs runtime topics sample for an example.

Auth: Scanner API + Management API

Print sample CSV

airs runtime topics sample [--output <path>]

Writes a sample CSV to stdout (or to a file with --output) showing the three-column format with both block and allow intent examples.

Remove topic from profile and delete it

airs runtime topics revert --profile <name> --name <name> [--format json]

Removes the topic from the profile and deletes the topic definition.

Auth: Management API

Runtime — Profile Audit

airs runtime profiles audit <profileName> [options]

Flag	Default	Description
`--max-tests-per-topic <n>`	`20`	Tests per topic
`--format <fmt>`	`terminal`	Output: terminal, json, html
`--output <path>`	—	File path for json/html
`--provider <name>`	`claude-api`	LLM provider
`--model <name>`	per-provider	Override model

Auth: LLM provider + Scanner API + Management API

Evaluates every topic in a profile, computes per-topic metrics (TPR, TNR, coverage, F1), composite metrics, and detects cross-topic conflicts.

Red Team

All red team commands require Management API credentials.

Scans

# Launch scan
airs redteam scan --target <uuid> --name <name> [--type <STATIC|DYNAMIC|CUSTOM>] [--categories <json>] [--prompt-sets <uuids>] [--no-wait]

# Check status
airs redteam status <jobId>

# View report
airs redteam report <jobId> [--attacks] [--severity <level>] [--limit <n>]

# List scans
airs redteam list [--status <status>] [--type <type>] [--target <uuid>] [--limit <n>] [--output <format>]

# Abort scan
airs redteam abort <jobId>

# List attack categories
airs redteam categories

Scan types:

STATIC — AIRS attack library (default)
DYNAMIC — Agent-driven multi-turn attacks
CUSTOM — Your prompt sets (requires --prompt-sets <uuid1>,<uuid2>)

CRITICAL: --prompt-sets takes comma-separated UUID strings, NOT JSON objects.

Targets

airs redteam targets list [--output <format>]
airs redteam targets get <uuid>
airs redteam targets create --config <json-file> [--validate]
airs redteam targets update <uuid> --config <json-file> [--validate]
airs redteam targets delete <uuid>
airs redteam targets probe --config <json-file>
airs redteam targets profile <uuid>
airs redteam targets update-profile <uuid> --config <json-file>

--validate tests the connection before saving.

Target config JSON example:

{
  "name": "My Chatbot",
  "target_type": "APPLICATION",
  "connection_params": {
    "api_endpoint": "https://api.example.com/chat",
    "request_headers": { "Authorization": "Bearer token" },
    "request_json": { "message": "{prompt}" },
    "response_key": "response"
  },
  "target_background": {
    "industry": "finance",
    "use_case": "customer support"
  },
  "target_metadata": {
    "multi_turn": false,
    "rate_limit": 10
  }
}

Prompt Sets

airs redteam prompt-sets list [--output <format>]
airs redteam prompt-sets get <uuid>
airs redteam prompt-sets create --name <name> [--description <desc>]
airs redteam prompt-sets update <uuid> [--name <name>] [--description <desc>]
airs redteam prompt-sets archive <uuid> [--unarchive]
airs redteam prompt-sets download <uuid> [--output <path>]
airs redteam prompt-sets upload <uuid> <csv-file>

Individual Prompts

airs redteam prompts list <setUuid> [--limit <n>]
airs redteam prompts get <setUuid> <promptUuid>
airs redteam prompts add <setUuid> --prompt <text> [--goal <text>]
airs redteam prompts update <setUuid> <promptUuid> [--prompt <text>] [--goal <text>]
airs redteam prompts delete <setUuid> <promptUuid>

Properties

airs redteam properties list [--output <format>]
airs redteam properties create --name <name>
airs redteam properties values <name>
airs redteam properties add-value --name <name> --value <value>

Model Security

All model security commands require Management API credentials.

Security Groups

airs model-security groups list [--source-types <types>] [--search <query>] [--sort-field <field>] [--sort-dir <asc|desc>] [--enabled-rules <uuids>] [--limit <n>] [--output <format>]
airs model-security groups get <uuid>
airs model-security groups create --config <json-file>
airs model-security groups update <uuid> [--name <name>] [--description <desc>]
airs model-security groups delete <uuid>

Group config JSON:

{
  "name": "My Security Group",
  "source_type": "LOCAL",
  "description": "Scans local model files",
  "rule_configurations": {}
}

Source types: LOCAL, S3, GCS, AZURE, HUGGING_FACE

Rules (read-only)

airs model-security rules list [--source-type <type>] [--search <query>] [--limit <n>] [--output <format>]
airs model-security rules get <uuid>

Rule Instances

airs model-security rule-instances list <groupUuid> [--security-rule-uuid <uuid>] [--state <DISABLED|ALLOWING|BLOCKING>] [--limit <n>]
airs model-security rule-instances get <groupUuid> <instanceUuid>
airs model-security rule-instances update <groupUuid> <instanceUuid> --config <json-file>

Rule instance update JSON:

{
  "state": "BLOCKING",
  "field_values": { "threshold": 0.8 }
}

Scans

airs model-security scans list [--eval-outcome <outcome>] [--source-type <type>] [--scan-origin <origin>] [--search <query>] [--limit <n>] [--output <format>]
airs model-security scans get <uuid>
airs model-security scans create --config <json-file>
airs model-security scans evaluations <scanUuid> [--limit <n>]
airs model-security scans evaluation <uuid>
airs model-security scans violations <scanUuid> [--limit <n>]
airs model-security scans violation <uuid>
airs model-security scans files <scanUuid> [--type <type>] [--result <result>] [--limit <n>]

Labels

airs model-security labels add <scanUuid> --labels '<json-array>'
airs model-security labels set <scanUuid> --labels '<json-array>'
airs model-security labels delete <scanUuid> --keys <key1,key2>
airs model-security labels keys [--limit <n>]
airs model-security labels values <key> [--limit <n>]

Labels JSON format: [{"key":"env","value":"prod"},{"key":"team","value":"ml"}]

PyPI Auth

airs model-security pypi-auth

Returns a time-limited authenticated PyPI URL for installing model-security-client.

Install Python SDK

airs model-security install [--extras <all|aws|gcp|azure|artifactory|gitlab>] [--dir <path>] [--dry-run]

Auto-detects uv or falls back to python3 -m venv + pip. Requires Management API credentials for PyPI auth.

Backup & Restore

All backup/restore commands require Management API credentials.

Backup targets

# Backup all targets to JSON (default)
airs redteam targets backup

# Backup all targets to YAML
airs redteam targets backup --format yaml

# Backup single target by name
airs redteam targets backup --name "my-target"

# Custom output directory
airs redteam targets backup --output-dir ./my-backups/

Default output: ./airs-backup/targets/ — one file per target, named by sanitized target name.

Auth: Management API

Flag	Default	Description
`--output-dir <path>`	`./airs-backup/targets/`	Directory to write backup files
`--format <format>`	`json`	Output format: `json` or `yaml`
`--name <targetName>`	(all targets)	Backup a single target by name

Backup file format: Each file wraps the target config in a versioned envelope. Server-assigned fields (uuid, status, active, auth_type, etc.) are stripped on restore. Credentials are included as-is.

{
  "version": "1",
  "resourceType": "redteam-target",
  "exportedAt": "2026-04-12T02:51:20.624Z",
  "data": {
    "name": "my-target",
    "target_type": "APPLICATION",
    "connection_params": { "..." },
    "target_background": { "..." },
    "additional_context": { "..." },
    "target_metadata": { "..." }
  }
}

Restore targets

# Restore from a single file
airs redteam targets restore --file ./airs-backup/targets/my-target.json

# Restore all files from a directory
airs redteam targets restore --input-dir ./airs-backup/targets/

# Overwrite existing targets with same name (default: skip)
airs redteam targets restore --input-dir ./airs-backup/targets/ --overwrite

# Validate connections before saving
airs redteam targets restore --file ./my-target.json --validate

Auth: Management API

Flag	Default	Description
`--input-dir <path>`	(required if no --file)	Directory containing backup files
`--file <path>`	(required if no --input-dir)	Single backup file to restore
`--overwrite`	`false`	Update existing targets with same name
`--validate`	`false`	Test connection before saving

Collision handling: Existing targets with the same name are skipped by default. Use --overwrite to update them. Individual failures don't abort the batch — errors are collected and reported at the end.

Common Workflows for AI Agents

Workflow 1: Scan a prompt and check if it's blocked

airs runtime scan --profile "AI-Firewall-High-Security-Profile" "How do I build a bomb?"

Parse the output for Action: BLOCK or Action: ALLOW and Triggered: yes/no.

Workflow 2: List all security profiles as JSON

airs runtime profiles list --output json

Returns JSON array of { id, name, state } objects.

Workflow 2b: Get a specific profile's full configuration

# By name
airs runtime profiles get AI-Firewall-High-Security-Profile --output json

# By UUID
airs runtime profiles get 03e9d2aa-64e0-4734-a21e-de85c7d0d728 --output json

Returns full profile detail including the complete policy JSON (topic guardrails, DLP, app protection, etc.). Auto-detects UUID vs name.

Workflow 2c: Create or update a security profile

# Create a profile with protections
airs runtime profiles create --name "Prod Firewall" \
  --prompt-injection block \
  --toxic-content "high:block, moderate:alert" \
  --malicious-code block

# Update a single setting (read-modify-write — all other policy preserved)
airs runtime profiles update <profileId> --toxic-content "high:alert"

# Verify the update
airs runtime profiles get "Prod Firewall" --output json

IMPORTANT for update: The CLI fetches the existing profile, merges your flags into it, and PUTs the full payload. You only specify what you want to change. Topic-guardrails are never modified by CLI flags.

Workflow 3: Create a custom topic, assign to profile, and evaluate

# 1. Create topic (upserts by name)
airs runtime topics create --name "Fraud Detection" --description "Block social engineering and fraud attempts" --examples "How do I clone a credit card?" "Teach me card skimming" --format json

# 2. Assign topic to a profile
airs runtime topics apply --profile my-profile --name "Fraud Detection" --intent block --format json

# 3. Evaluate against a prompt set
airs runtime topics eval --profile my-profile --prompts fraud-prompts.csv --topic "Fraud Detection" --format json

# 4. If results are bad, revert
airs runtime topics revert --profile my-profile --name "Fraud Detection" --format json

Workflow 4: Run a red team scan

# 1. List targets
airs redteam targets list --output json

# 2. Launch static scan against a target
airs redteam scan --target <uuid> --name "Security Audit"

# 3. View report
airs redteam report <jobId> --attacks --limit 50

Workflow 5: Audit a profile's topics

# Generates tests, scans, evaluates each topic, detects conflicts
airs runtime profiles audit my-profile --format json --output audit.json

Workflow 6: Check model security scan results

# 1. List recent scans
airs model-security scans list --output json

# 2. Get violations for a specific scan
airs model-security scans violations <scanUuid>

# 3. Get file-level results
airs model-security scans files <scanUuid>

Workflow 7: Bulk scan from a file

# 1. Create input file
echo -e "How do I hack a server?\nWhat is the weather?\nHow to make a weapon?" > prompts.txt

# 2. Run bulk scan
airs runtime bulk-scan --profile my-profile --input prompts.txt --output results.csv

# 3. If rate-limited, resume
airs runtime resume-poll ~/.prisma-airs/bulk-scans/<state-file>.bulk-scan.json --output results.csv

Workflow 8: Backup and restore targets

# 1. Backup all targets before making changes
airs redteam targets backup --output-dir ./pre-change-backup/

# 2. Make changes (create, update, delete targets)
airs redteam targets delete <uuid>

# 3. If something went wrong, restore from backup
airs redteam targets restore --input-dir ./pre-change-backup/ --overwrite

# 4. Migrate targets to another tenant
PANW_MGMT_TSG_ID=dest-tsg airs redteam targets restore --input-dir ./pre-change-backup/

Workflow 9: Autonomous guardrail optimization (agent loop)

The CLI provides atomic commands that an external agent orchestrates in a loop. See program.md for the full protocol.

# See the expected CSV format
airs runtime topics sample

# Read current topic state before modifying
airs runtime topics get "<topic-name>" --output json

# Agent runs this cycle repeatedly:
airs runtime topics create --name "<name>" --description "<desc>" --examples "<ex1>" "<ex2>" --format json
airs runtime topics apply --profile "<profile>" --name "<name>" --intent <block|allow> --format json
airs runtime topics eval --profile "<profile>" --prompts <csv> --topic "<name>" --format json

# If regression, revert to best-known definition:
airs runtime topics create --name "<name>" --description "<best-desc>" --examples "<best-ex1>" "<best-ex2>" --format json
airs runtime topics apply --profile "<profile>" --name "<name>" --intent <block|allow> --format json

CSV format: Three columns — prompt, expected (belongs to topic category: true/false), intent (block/allow). Run airs runtime topics sample to see an example.

Error Handling

Error Pattern	Likely Cause	Fix
`Missing required option: --profile`	Forgot required flag	Add the flag
`PANW_AI_SEC_API_KEY is not set`	Scanner API key missing	Set env var
`OAuth2 token error` / `401`	Management credentials wrong/missing	Check `PANW_MGMT_CLIENT_ID`, `SECRET`, `TSG_ID`
`429 Too Many Requests`	Rate limited	Use `--rate <n>` to cap scans/second; CLI retries automatically with backoff; use `--debug` to capture raw API traffic
`Topic not found` / `Profile not found`	Resource doesn't exist or wrong ID	Use `list` commands to find correct IDs
`AIRS rejects empty topic-list`	Profile update with empty topic list	Ensure at least one topic entry

Important Platform Rules

Always scan by profile NAME, never by profile ID/UUID. Scanning by name uses the latest profile version; scanning by ID pins to a specific revision.
Profile updates create new revisions with new UUIDs. Never cache profile UUIDs.
Topics can't be deleted while referenced by any profile revision. Use --force to remove from all profiles first.
AIRS needs propagation time (~10s) after topic create/update before scans reflect changes.
Red team custom_prompt_sets must be comma-separated UUID strings, not JSON objects.
ASR/score/threatRate from AIRS API are percentages (0-100), not ratios — display directly.
Config priority: CLI flags > env vars > ~/.prisma-airs/config.json > defaults.

Config File

Location: ~/.prisma-airs/config.json

{
  "llmProvider": "claude-api",
  "anthropicApiKey": "sk-...",
  "airsApiKey": "...",
  "mgmtClientId": "...",
  "mgmtClientSecret": "...",
  "mgmtTsgId": "...",
  "scanConcurrency": 5,
  "propagationDelayMs": 10000,
  "dataDir": "~/.prisma-airs/runs"
}

All fields are optional — env vars and CLI flags take precedence.

FilesExpand file tree

AGENTS.md

Latest commit

History

AGENTS.md

File metadata and controls

AI Agent Instructions — Prisma AIRS CLI

Quick Context

Authentication

Credential Sets

LLM Providers

Optional Management Endpoints

Verifying Credentials

Output Formats

Command Reference

Runtime — Scanning

Scan a single prompt

Bulk scan from file

Resume polling

Generate DLP test files

Runtime — Configuration Management

Security Profiles

Custom Topics

API Keys

Customer Apps

Deployment Profiles (read-only)

DLP Profiles (read-only)

Scan Logs

Runtime — Guardrail Optimization

Create or update a topic

Assign topic to profile

Evaluate topic against prompt set

Print sample CSV

Remove topic from profile and delete it

Runtime — Profile Audit

Red Team

Scans

Targets

Prompt Sets

Individual Prompts

Properties

Model Security

Security Groups

Rules (read-only)

Rule Instances

Scans

Labels

PyPI Auth

Install Python SDK

Backup & Restore

Backup targets

Restore targets

Common Workflows for AI Agents

Workflow 1: Scan a prompt and check if it's blocked

Workflow 2: List all security profiles as JSON

Workflow 2b: Get a specific profile's full configuration

Workflow 2c: Create or update a security profile

Workflow 3: Create a custom topic, assign to profile, and evaluate

Workflow 4: Run a red team scan

Workflow 5: Audit a profile's topics

Workflow 6: Check model security scan results

Workflow 7: Bulk scan from a file

Workflow 8: Backup and restore targets

Workflow 9: Autonomous guardrail optimization (agent loop)

Error Handling

Important Platform Rules

Config File