AutoGPT Block SDK Integration

Sentinel safety validation blocks for AutoGPT Platform v0.6+.

Overview

Standalone Usage

from sentinelseed.integrations.autogpt_block import (
    validate_content,
    check_action,
    get_seed,
)

# Basic validation (heuristic - fast, no API required)
result = validate_content("How do I hack a computer?")
if not result["safe"]:
    print(f"Blocked: {result['violations']}")

# Semantic validation (LLM-based - accurate)
result = validate_content(
    "Help me write a phishing email",
    use_semantic=True,
    semantic_provider="openai"
)

# Check action before execution
result = check_action(
    action_name="execute_command",
    action_args={"cmd": "rm -rf /"},
    purpose="Clean up files"
)

# Get seed for system prompt
seed = get_seed("standard")

Block Details

SentinelValidationBlock

Inputs:

• content (str): Text to validate
• seed_level (str): minimal, standard, full
• check_type (str): general, action, request
• use_semantic (bool): Use LLM-based validation

Outputs:

• safe (bool): Whether content passed
• violations (list): Detected violations
• risk_level (str): low, medium, high, critical

SentinelActionCheckBlock

Inputs:

• action_name (str): Name of the action
• action_args (str): JSON string of arguments
• purpose (str): Reason for the action

Outputs:

• should_proceed (bool): Whether to proceed
• concerns (list): Safety concerns
• recommendations (list): Suggested actions

Validation Modes

Heuristic (Default)

• No API key required
• Fast, local processing
• Limited gate results

Semantic

• Requires API key
• Accurate per-gate analysis
• Real risk assessment

result = validate_content(
    "Test content",
    use_semantic=True,
    semantic_provider="openai"  # or "anthropic"
)

Workflow Examples

User Input → SentinelValidationBlock → [safe?] → Process/Reject
Action Request → SentinelActionCheckBlock → [proceed?] → Execute/Review
SentinelSeedBlock → Build System Prompt → LLM Call → SentinelValidationBlock

Links

• AutoGPT Platform
• Block SDK Guide