Security Model

I-Lang is an instruction format, not an authorization bypass. All dangerous operations require host runtime confirmation, permission checks, and audit logging.

Core principle

I-Lang is an instruction format, not an authorization bypass.

I-Lang structures what an agent is told to do. It does not grant the agent permission to do it. All dangerous operations require host runtime confirmation, permission checks, and audit logging.

Verb risk classification

Safe verbs (no side effects)

These verbs read, analyze, or format data without modifying external state.

VerbRiskReason
READ, GET, LIST, STRMSafeRead-only access
FMT, CONV, SPLIT, MERGE, MAP, FILT, SORT, DEDUSafeIn-memory transformation
SCAN, MTCH, CNT, STAT, EVAL, SCOR, RANKSafeAnalysis, no state change
OUT, DISP, EXPT, PRNT, LOGSafeOutput to user
HELP, DESC, INTR, NOOPSafeInformational

Guarded verbs (require confirmation)

These verbs modify files, send data, or change external state. Host runtime must require user confirmation before execution.

VerbRiskRequired safeguard
WRITGuardedConfirm file path and content before writing
MOVEGuardedConfirm source and destination
COPYGuardedConfirm destination does not overwrite
SENDGuardedConfirm recipient and content
SYNCGuardedConfirm sync direction and scope
CACHGuardedConfirm cache storage location
SAVEGuardedConfirm state checkpoint location

Dangerous verbs (require explicit authorization)

These verbs can cause data loss, execute arbitrary code, or modify production systems. Host runtime must enforce permission checks, dry-run mode, sandbox execution, and audit logging.

VerbRiskRequired safeguards
DELDangerousUser confirmation + dry-run preview + audit log
RUNDangerousSandbox execution + command preview + user approval
DPLODangerousTarget environment confirmation + rollback plan + audit log

Runtime requirements

Any runtime that executes I-Lang instructions must implement these safeguards:

RequirementDescription
Permission checkVerify the agent has authorization for the target resource before executing guarded or dangerous verbs
User confirmationPresent a preview of what will happen and require explicit user approval for guarded and dangerous verbs
Dry-run modeAllow users to preview the effect of a chain without executing it
Sandbox executionDangerous verbs like RUN must execute in an isolated environment by default
Audit logAll guarded and dangerous verb executions must be logged with timestamp, verb, target, modifier, and user who approved
Host policy overrideThe host system can restrict or block any verb regardless of I-Lang instructions

What I-Lang does NOT do

I-Lang does NOTExplanation
Grant permissionsI-Lang structures intent. The host runtime enforces permissions.
Bypass safety layersBehavioral definitions (::GENE{}) operate within the model's existing safety constraints.
Execute code directly[RUN|cmd=...] is an instruction to the runtime, not direct code execution.
Override host policiesIf the host blocks DEL, no I-Lang instruction can override that.

Reporting security issues

If you discover a security concern with the I-Lang specification or any I-Lang tool, report it via GitHub security advisory.