Policies

Agentic Runtime policies decide what the agent does with each kind of detection: observe it, mask it or block it. You set them for the whole organization and refine them for teams, members and devices.

Policy Cascade

TEXT

Organization  →  Team  →  Member  →  Device

Organization  →  Team  →  Member  →  Device

Scope	Screen	Use it for
Organization	Organization	The default rules applied to every device
Team	Teams	Rules for a whole team at once, such as stricter blocking for juniors or enforce-always for compliance
Member	Users	Exceptions for one person
Device	Devices	Exceptions for one machine

A more specific scope overrides the broader one. When a member is in several teams, the strictest policy wins on each axis.

Each device shows its effective policy and where every setting comes from, so you can always see why something was blocked or allowed.

Configure Detections

Open Agentic runtime → Organization and click New rule, or edit a team, member or device policy.

ByteHide Agentic Runtime Configure detections drawer with credential leaks, prompt injection, dangerous tools, malicious MCPs, risky skills and LLM models Click to expand

Detection	Modes	Extra settings
Credential leaks	Observe, Mask, Block	—
Prompt injection	Off, Block	—
Dangerous tools	Observe, Block	Tune each command in Inventory → Tools
Malicious MCPs	Off, Block	Block specific MCP servers by name. Manage servers in Inventory → MCPs
Risky skills	Off, Block	Block specific skills by name
LLM models	Off, Block	Block specific models, or entire AI providers however they are reached (direct API, Bedrock, OpenRouter, OpenAI-compatible gateways)

Names you add to a block list are hard vetoes: they always block when matched, whatever the mode.

Presets

Preset	Behavior
Strict	Blocks every detection family. A good baseline for junior developers
Standard	A balance of blocking and observing
Relaxed	Observes only: logs, never blocks. A good baseline for senior developers

Configure by Voice

Click Configure with AI, or Start conversation on the voice banner, and describe what you want in your own words. The assistant drafts a policy proposal; you review it and apply it or discard it. Nothing changes automatically.

Custom Rules

Agentic runtime → Custom Rules adds your own detection patterns, run alongside the built-in catalog (AWS, GitHub and Stripe keys, prompt injections...). Custom rules are additive: the built-in catalog keeps running.

Patterns use RE2 syntax: matching runs in linear time with no risk of catastrophic backtracking, but lookarounds and backreferences are not supported.

Repository Path Rules

Agentic runtime → Repo lists every git repository a device in your fleet has worked in. Add path-deny rules per scope (organization, team, member or device): the agent blocks any read, edit, write, glob or grep tool call against a denied path before the file content reaches the model.

Developers can also commit a .aiblock file at the repository root. Its rules merge with the ones you set in the console, and the agent discovers it on its next start.