Security Analysis

The security middleware assesses tool call risk and optionally requires user confirmation before executing dangerous operations.

Risk Levels

use synaptic::middleware::RiskLevel;

pub enum RiskLevel {
    None,
    Low,
    Medium,
    High,
    Critical,
}

SecurityAnalyzer Trait

Assesses the risk level of a tool call based on its name and arguments.

use synaptic::middleware::SecurityAnalyzer;

#[async_trait]
pub trait SecurityAnalyzer: Send + Sync {
    async fn assess(&self, tool_name: &str, args: &Value) -> Result<RiskLevel, SynapticError>;
}

RuleBasedAnalyzer

Maps tool names and argument patterns to risk levels.

use synaptic::middleware::{RuleBasedAnalyzer, RiskLevel};

let analyzer = RuleBasedAnalyzer::new()
    .with_default_risk(RiskLevel::Low)
    .with_tool_risk("delete_file", RiskLevel::High)
    .with_tool_risk("read_file", RiskLevel::None)
    .with_arg_pattern("path", "/etc", RiskLevel::Critical);

Argument patterns elevate the risk when a tool argument value contains the specified substring.

ConfirmationPolicy

Determines whether a tool call at a given risk level requires user confirmation.

use synaptic::middleware::{ThresholdConfirmationPolicy, RiskLevel};

// Require confirmation for High and Critical risk
let policy = ThresholdConfirmationPolicy::new(RiskLevel::High);

SecurityConfirmationCallback

Implement this trait to define how confirmation is obtained from the user.

use synaptic::middleware::{SecurityConfirmationCallback, RiskLevel};

struct CliConfirmation;

#[async_trait]
impl SecurityConfirmationCallback for CliConfirmation {
    async fn confirm(
        &self,
        tool_name: &str,
        args: &Value,
        risk: RiskLevel,
    ) -> Result<bool, SynapticError> {
        println!("Tool '{}' has {:?} risk. Allow? [y/N]", tool_name, risk);
        // read user input...
        Ok(true)
    }
}

SecurityMiddleware

Combines the analyzer, policy, and callback into a single middleware.

use synaptic::middleware::SecurityMiddleware;

let middleware = SecurityMiddleware::new(
    Arc::new(analyzer),
    Arc::new(policy),
    Arc::new(CliConfirmation),
)
.with_bypass(["get_weather"]);  // these tools skip security checks

let options = AgentOptions {
    middleware: vec![Arc::new(middleware)],
    ..Default::default()
};

When a tool call is intercepted, the middleware assesses its risk, checks the policy, and if confirmation is required, invokes the callback. If the user rejects, the tool call returns an error.

Keyboard shortcuts