What happens when a research team teaches a compiler to turn security wishes into runnable network maps?

Enterprise networks start as words.

Compliance officers write policies. Architects draw zones. Engineers configure firewalls. The gap between the first sentence and the last command is where mistakes hide. TopoIntent closes that gap with a compiler pipeline that ingests natural-language intent, retrieves reference architectures from a curated library of 22 templates, aligns the intent to the template through staged fusion, checks the resulting topology against CIS Controls v8.1.2 safeguards visible at the topology layer, repairs structural gaps with additive schema-preserving edits, and exports Mininet scripts with kernel-level iptables ACLs for executable reachability and allow/deny tests.

The template library covers five scenarios. The retrieval set contains 44 synthetic intents. A held-out set of 7 templates and 14 intents comes from finance and government scenarios excluded from retrieval. On that held-out set, additive repair lifts topology-visible CIS satisfaction from 0.78 to 1.00 in fewer than 1.5 rounds on average. One feedback round raises the post-ACL policy pass rate from 0.78 to 0.88.

How the pipeline constrains generation

A schema contract defines what a valid topology looks like. Zones, boundary devices, inter-zone paths, and access-control policies must conform. The contract acts as a guardrail during generation. Anything that violates the schema is rejected before it reaches the compliance checker.

Dense-vector search fetches the closest reference architectures from the curated template library. The search operates on embeddings of the intent text. Retrieved templates provide structural priors — zone layouts, typical boundary devices, common path patterns — that the fusion stage adapts to the specific intent.

Staged fusion performs two jobs. First, intent-template alignment maps the underspecified natural language onto the template structure. Second, security completion adds missing controls implied by the intent but absent from the template. The paper does not detail the fusion algorithm beyond naming the stages.

Compliance checking at the topology layer

CIS Controls v8.1.2 safeguards are evaluated only where they are visible in the topology. Network segmentation, boundary protection, and access-control enforcement appear in the generated structure. Safeguards that depend on host configuration, logging, or procedural controls remain outside the check. Unresolved cases are marked for manual review rather than assumed compliant.

Additive schema-preserving edits repair structural gaps. The edits add zones, devices, or paths without removing existing elements or violating the schema. The paper reports that repair converges in fewer than 1.5 rounds on average for the held-out set. Each round re-evaluates the topology against the CIS safeguards.

From topology to executable testbed

The final artifact is a Mininet script. Kernel-level iptables ACLs implement the access-control policies. The script spins up a virtual network where reachability and allow/deny tests run as executable code. This moves verification from static analysis to runtime behavior.

No public benchmark existed for the requirement-to-topology task. The authors constructed their evaluation set from reference security architecture diagrams. The retrieval set and held-out set are disjoint by scenario. Finance and government scenarios appear only in the held-out set. The paper does not state how many real-world intents were tested versus synthetic ones.

The system marks unresolved cases for manual review. It does not auto-approve them. The feedback loop that raises the post-ACL policy pass rate from 0.78 to 0.88 operates on the held-out set. The paper does not describe the feedback mechanism beyond counting one round.

A compiler approach can bridge intent and executable topology. The pipeline — schema contract, template retrieval, staged fusion, compliance checking, additive repair, Mininet export — runs end to end. The numbers come from a constructed benchmark. How the system behaves on production intents with ambiguous language, conflicting requirements, or legacy constraints remains unmeasured.