Protected by the 3S framework
Observe, understand, and intervene across every request flowing through your agentic application.
dFenceRuntime security
Your agentic CISO,
always on.
dFence is a runtime AI guardrails platform for agentic applications. It observes every request, understands policy defiance, and intervenes with the right action when prompts, tools, data, or agent behavior become unsafe.
dFence at a glance
- Agent request
Prompts, context, data, and tool calls
- Intent and policy evaluation
Compare the action with its intended purpose
- Runtime decision
Allow, block, redact, constrain, or alert
dFence capabilities
Protection that understands, then acts.
Intent-based access
Allow tools and MCP calls only when the live action matches the agent’s declared happy path.
AI policy writer
Translate vulnerabilities and business boundaries into enforceable policies.
Controls matched to criticality
Use more than 100 checks with policies deployed as enforcers or observers based on criticality—avoiding heavy, one-size-fits-all guardrails.
How it works
Observe. Understand. Intervene.
dFence creates a closed decision loop around every agent request, from raw telemetry to autonomous action.
dFence workflow
- Observe
Every request, response, data source, and tool call.
- Understand
Intent, context, policy defiance, and business risk.
- Intervene
Block, redact, constrain, alert, or allow autonomously.
Illustrative workflow
Access control for what the agent is trying to do.
A support agent has access to customer records, but its approved purpose is billing support and account summaries. It attempts an unrelated bulk customer export.
Illustrative example
- Declared happy path
Billing support and account summaries within defined data boundaries.
- Intent check
dFence compares the request, context, MCP call, and action parameters with the expected intent.
- Policy decision
The bulk export conflicts with the approved purpose. The tool call is blocked and an audit event is created.
Trusted low-risk calls can proceed. Untrusted, sensitive, or mismatched actions are blocked, constrained, redacted, or sent for approval.
Evidence and integrations
Connect policy decisions with evidence.
Audit trails
Maintain audit trails for every request to understand the agent action and the runtime response.
Framework mapping
Map findings against OWASP Top 10 for LLMs, MITRE ATLAS, NIST, the EU AI Act, and other relevant frameworks.
Part of the Dwaar platform