Secure
PydanticAI
PydanticAI security requires type-safe runtime governance that validates agent tool calls against execution boundaries without breaking the developer workflow. SupraWall shims the execution context of PydanticAI agents to provide real-time interception, enabling teams to enforce security policies and audit autonomous behavior at the runtime level.
Type Safety is not Execution Safety
PydanticAI is praised for its type safety, but type safety is not security. A well-typed `delete_user(user_id: str)` call is still destructive if the agent decides to trigger it maliciously or via indirect prompt injection. SupraWall provides the governance layer that sits between PydanticAI's internal tool dispatcher and your environment.
Quick Implementation
# Install the official integration package
pip install suprawall-python
# Secure your agent executor
from pydantic_ai import Agent
from suprawall.pydantic_ai import Guard
my_agent = Agent(...)
# Wrap it for runtime interception
secured_agent = Guard(my_agent, policy_id="finance_v1")
Injection Blocking
SupraWall monitors PydanticAI's tool calling patterns specifically looking for indirect prompt injection vectors in RAG-provided data.
Type-Safe Audits
Because PydanticAI uses strongly-typed models, SupraWall logs audit data with full JSON schema compliance, making it easy to analyze failures.
EU AI Act Compliance
For developers using PydanticAI in financial or regulated sectors, SupraWall fulfills the requirement for Deterministic Governance (Article 14). We ensure that every 'high-risk' tool call is checked against a non-LLM policy engine before execution.
Deterministic Policy Enforcement
In production, an agent shouldn't be allowed to execute tools like db_execute or send_api_request without a boundary verification. SupraWall's integration follows the AGPS Spec to intercept these calls before they hit your infrastructure.
Native Features:
- Real-time interception of 'FunctionTool' calls.
- Validation against runtime budget caps to prevent $500 billing surprises.
- Automatic 'Fail-Closed' behavior if the security shim is disconnected.
- Seamless integration with PydanticAI's 'RunContext' for dependency injection.
Internal Report
Stopping Infinite Loops
How to prevent recursive tool calls in your PydanticAI swarm.
Theoretical Base
Understanding ARS
The missing layer between your LLM and your environment.