Securing Retrieval-Augmented Code Generation via Contextual Knowledge Injection: A Case for Embedded IoT Applications

Repository-grounded retrieval-augmented code generation (RACG) is increasingly used in embedded IoT development by retrieving code and documentation from a pinned RTOS/SDK repository (e.g., Zephyr OS). In this setting, security risks are often version-inherited: even without retrieval poisoning, generated applications may invoke benign-looking public APIs that transitively reach vulnerable internal routines in the pinned snapshot, thereby inheriting known CVEs. Existing secure RACG pipelines largely focus on task-level intent and generic vulnerability patterns, which can miss repository- and version-specific exposure. Meanwhile, conventional CVE scanners can flag vulnerable locations but cannot determine whether those vulnerabilities are reachable through the public APIs that the generator commits to during repository-grounded generation. In this paper, we present IoTRAGuarder, a contextual knowledge injection framework that aligns security hardening with generation-time API selection under repository grounding. IoTRAGuarder (i) recovers auditable reverse call chains from CVE-localized internals to exposing public APIs via static analysis plus an evidence-gated LLM to bridge indirections and macro-driven "call-graph islands", (ii) constructs a version-aware security knowledge base that binds affected version intervals to exposed public APIs with prompt-ready constraints, safer alternatives, or avoidance/upgrade guidance, and (iii) performs dual-layer, API-aligned online retrieval to inject concise, version-matched constraints into the final prompt. We evaluate IoTRAGuarder on 44 real-world Zephyr tasks across four LLMs. Compared to the prior state-of-the-art secure RACG baseline, IoTRAGuarder improves the overall security success rate from 5.11% to 78.41%.