Cortega Agentic Firewall

Protect agent conversations
and tool calls using smart application sensing technology.

Block prompt injection, social engineering, poisoned context and policy violations. Sense and manage application level threat levels automatically using agents. Apply basic and agentic guardrails to model traffic and eligible tool calls.

Explore the attack examples →

Basic guardrails

Inspect content.

  • PII and secretsDetect sensitive fields; mask or reject under the configured policy.
  • Unsafe contentUse moderation providers, local models, patterns and custom checks.
  • Prompt and response rulesApply request checks before the model and response checks before delivery.

Agentic guardrails

Inspect intent and actions.

  • Social engineering and fraudCheck authority claims, false urgency, pretexts and incremental escalation.
  • Conversation risk and injectionRead the conversation with its roles and inspect eligible tool content.
  • Policy and tool-call reviewCheck proposed calls against allowed tools, arguments and destinations; review answers against approved policy.

Attacks and controls

See the behavior each control addresses.

Social engineering detection

Detect impersonation, false urgency, and pretexts that manipulate an agent.

Impersonated caller Agent conversation Identity and intent checks

Explanatory attack path

Try the agent ↗

Financial fraud prevention

Detect account-takeover, identity-failure, and pressure tactics that enable fraud.

Fraud pretext Account workflow Fraud policy

Explanatory attack path

Try the agent ↗

Policy compliance

Enforce your policies. Explore the refund-policy demo.

Refund request Support agent Approved policy

Explanatory attack path

Try the agent ↗

Agent Poisoning

Detect malicious tool-result instructions that redirect an agent’s behavior.

Poisoned tool result Application agent Content inspection

Explanatory attack path

Try the agent ↗

Data exfiltration & PII

Detect sensitive data and credentials. Mask or block disclosure under your policy.

Sensitive data Agent response Mask or block

Explanatory attack path

Try the agent ↗

SIM Swap Detection

Detect identity-check bypasses and phone-transfer manipulation.

Phone-transfer request Support agent Identity check

Explanatory attack path

Try the agent ↗

Unauthorized vs Authorized Data Access

Compare permitted requests with attempts to access data without authorization.

Caller identity MCP tool call Access policy

Explanatory attack path

Try the agent ↗

Prompt Injection

Detect instructions that bypass rules or redirect the task.

Injected instruction Application agent Instruction check

Explanatory attack path

Try the agent ↗

Where enforcement happens

Two traffic paths. One policy gateway.

The application sends model calls and MCP calls through Cortega. Checks run on the applicable request and response paths; the model and tools remain separate destinations.

Application agent sends LLM calls and MCP calls through Cortega to separate model providers and MCP tools, with request and response checks on both paths
Explanatory topology. Only traffic routed through Cortega and covered by an active policy is inspected.
Policy scope and operating limits

Bind checks to the relevant teams, models, applications, MCP servers or tools. Available actions and streaming behavior depend on the guardrail. Tool-call review checks a proposed call; it does not run the tool itself. Application threat policies can make configured rules stricter when the application’s threat level rises.

Console

Configure a check. Inspect its decision.

Cortega guardrails configuration
Actual Cortega console screenshot · Select to open the complete image.

Measured model safety

Attack success and harmless refusals.

October 2026 · Ten models · Historical in-process guardrail evaluation. This is not an agent penetration-test result.

Inspect the complete benchmark and data →

Review the policies for your model and tool traffic.

Book a demo →