Agent & AI Security
Apply policy to tool calls, data access, and handoffs before execution. Let routine actions proceed and hold high-stakes actions for review.
Build across runtimes
Govern in one place
Create, share, and run agents in Agent Hub. First Recon orchestrates cloud agents across OpenAI, Anthropic, and Copilot Studio, with company knowledge, tools, and policies in one layer.
Scope the tools an agent can use
Agents join the governed path through the gateway and MCP or OAuth tool connections. Each call runs within the permission grant it was given.
Keep the policy with the action
Human to model, agent to tool, agent to agent: identity, data, and destination stay attached as work moves between them.
It takes AI
to secure AI
A valid tool call can still carry the wrong intent. Semantic Security evaluates what the agent is trying to do against the data, destination, and policy.
Security Context Graph links behavior across sessions, identities, and data sources. Governed interactions and connected company data enrich threat and sensitive-data detection.
Explore the context graphKnow what happened
and why
Each governed action leaves a sealed record: who the agent acted for, what it attempted, which policy applied, and how the decision landed.
Stream metadata-only decisions to your SIEM. When a held action is approved or denied, the reviewer and outcome become part of the trail.
Explore Compliance & AuditControl the action
not just the prompt
What counts as an agent action?
A model call, tool call, or handoff to another agent on the governed path. This includes reading records, writing files, sending messages, and calling APIs. Each action is evaluated with identity and policy before it proceeds.
Which actions wait for a person?
Only those your policy routes to human review. Routine actions are evaluated inline. A high-stakes action holds before execution until a designated reviewer approves or denies it. Nothing executes while it waits.
Can a reviewer see the held action?
Yes. Hold for human review is the deliberate exception to metadata-only evidence. Designated reviewers see the held action, context, and triggering policy while it waits. After the decision, the trail retains who decided and the outcome.
Can policies redact sensitive data too?
Yes. Governed interactions can be allowed, redacted, held for human review, or blocked according to the configured policy. Tool permissions and action-level controls still apply.
- Detect
- Redact
- Hold
- Block