beam_security

beam_security checks content and actions against a rule base and returns a decision.

beam_security = { path = "../riverbeam/crates/beam_security" }

It depends on river_reed with the grl-rule-engine feature.

Item	Role
SecurityManagement	Trait: check(&SecurityRequest) -> SecurityDecision, enforce()
SecurityManager	new(policy), with_enforce(bool), is_ready()
DisabledSecurityManager	Allows everything; the pipeline default
SecurityRequest	new(stage), with_payload(map), with_context(SecurityContext)
SecurityContext	agent_id, tenant_id, user_id, interaction_id
SecurityDecision	action, severity, score, matched_rules, narration, user_message, violation_type; denies(), requires_hitl()
SecurityStage	BeforeInput, BeforeContextInjection, BeforeLlm, AfterLlm, BeforeTool, AfterTool, BeforeMemoryWrite, BeforeOutput, BeforeHitlResume, Runtime
SecurityAction	Allow, Deny, Warn, Sanitize, RequireApproval, Quarantine, Mask, Terminate
SecurityPolicy, ToolPolicy, HitlConfig	Tool lists, approval settings, runtime limits
PatternLoader	The regex scorers
SecurityEngine, verdict_from_narrations, RuleNarration	Lower-level engine access

BeforeLlm, AfterTool and BeforeHitlResume are always allowed (PASSTHROUGH_STAGES).

There are no feature flags.

The shipped rules and their current limitations are in Security.