Skip to content

Agent Shield

Agent Shield provides network policy enforcement and security monitoring for Agent Computers.

Every Agent Computer is isolated — dedicated kernel, separate memory/disk/network namespace. Network egress is filtered per-organization policy.

Network policies control which external domains an Agent Computer can reach. Policies are set at the organization level and apply to all computers in the organization.

Configure at: Settings > Security > Network Policy

ModeBehavior
Package Managers Only (default)Allows AI provider APIs, GitHub, and package registries. Blocks everything else.
All DomainsNo egress restrictions. The agent can reach any domain.
NoneBlocks all external domains. Add specific domains manually.

Default allowlist (Package Managers Only mode):

CategoryDomains
AI servicesapi.anthropic.com, api.openai.com, generativelanguage.googleapis.com
Code hostinggithub.com
NPMregistry.npmjs.org, npmjs.com
Pythonpypi.org, files.pythonhosted.org
Rustcrates.io, index.crates.io
Systemarchive.ubuntu.com, security.ubuntu.com

In Package Managers Only and None modes, you can add additional domains your agents need to access — internal APIs, specific SaaS services, or any other endpoint required by your workflows.


Shield monitors every network connection made by agents running inside Agent Computers.

  • Process-attributed connections — every TCP connect is tagged with the process name and full process lineage
  • DNS resolution — domain names are correlated to IP connections, so you see api.github.com instead of raw IPs
  • Byte accounting — total bytes sent and received per connection
  • Connection lifecycle — connect attempts, results, and close summaries

Shield is version-managed by the Rebyte platform. When monitoring is enabled for your organization:

  • Shield is automatically installed on every new computer at creation time
  • Existing computers receive Shield on their next resume
  • Version updates are applied automatically

The security dashboard provides real-time visibility into agent network activity across your organization.

Each computer is assigned a risk score based on its network behavior. Computers with unusual connection patterns — high volumes of outbound data, connections to uncategorized domains, or access to sensitive services — are flagged for review.

CategoryExamples
TrustedYour explicitly allowed domains
ProviderAI provider APIs (Anthropic, OpenAI, Google)
BenignPackage registries, documentation sites
SuspiciousUncategorized or unusual destinations
StateMeaning
NewAutomatically generated, not yet reviewed
OpenAcknowledged, under investigation
TriagedAssessed, assigned a priority
MutedSuppressed (known false positive or accepted risk)
ResolvedAddressed, no further action needed

The dashboard shows which computers in your organization have Shield installed and which do not, so you can verify that monitoring coverage is complete.


  • Audit trail — every connection is recorded with process attribution, timestamps, byte counts.
  • Data exfiltration prevention — domain allowlisting with default-deny mode.
  • Incident investigation — search exact connections by computer and time window. Process lineage shows which subprocess initiated each connection.
  • Compliance — SOC 2, ISO 27001, internal security policies.