Skip to main content
Security in Cascade is the process of identifying and containing adversarial attempts to manipulate, compromise, or abuse AI agents during execution. Rather than relying on traditional application security controls, Cascade uses purpose-built detection systems and runtime analysis to identify threats that are unique to agentic systems, such as prompt injection, data exfiltration, and alignment manipulation. This approach gives teams active defense against attacks that target agent reasoning, tool usage, and decision-making.

​
How security works in Cascade

Security in Cascade is detection-driven and enforcement-aware. As an agent executes, Cascade’s DeepStream system captures detailed traces of behavior. These traces are evaluated in real time against security-specific detection models. When a threat is identified, Cascade records the event and applies a response based on the configured enforcement mode. Security detection operates independently but integrates seamlessly with Cascade’s observability and safety systems.

​
Security capabilities

Cascade provides three primary security capabilities designed to address adversarial threats.

Threat Detection

Identify prompt injection, data exfiltration, and alignment manipulation in real time

Risk Map

Visualize security risk propagation across agents, tools, and data flows

Enforcement Control

Apply observe, detect, or enforce modes to security threats independently

​
Threat categories

Security threats are grouped into distinct categories based on attack vector and impact.
  • Prompt Injection: Attempts to override or manipulate agent instructions to produce unintended behavior
  • Data Exfiltration: Attempts to extract sensitive or unintended data through agent behavior or tool usage
  • Alignment Manipulation: Gradual or coordinated attempts to shift agent behavior or reasoning over time
Each category is detected using specialized models that analyze execution traces, reasoning patterns, and behavioral shifts.

​
Runtime evaluation

Security evaluation happens continuously while an agent runs. As traces are generated, Cascade:
  • Analyzes inputs for instruction manipulation
  • Monitors tool usage for unauthorized access patterns
  • Tracks behavioral shifts across historical executions
  • Evaluates outputs for data leakage indicators
  • Produces security findings tied to the trace
This enables security decisions to be made using full execution context rather than static rules or post-execution analysis. This allows teams to observe safety policy violations while actively blocking security threats.

​
Who security is for

The Security system is designed for:
  • Security teams monitoring adversarial risk
  • Platform teams deploying agents in production
  • Compliance teams managing threat exposure
  • Developers building agent workflows

​
Next steps

Threat Detection

Learn how Cascade detects and contains adversarial attacks

Risk Map

Understand how security risk propagates across your system