Safety Context Injection: Inference-time safety alignment based on static filtering and agent analysis
Safety Context Injection: Inference-Time Safety Alignment via Static Filtering and Agentic Analysis
AI summary
The researchers propose an inference-time safety alignment framework, Safety Context Injection, which separates safety evaluation from task generation and injects a structured risk report upfront. The lightweight single-pass filtering SMF has low latency, and the agent-loop-based DAF is more effective against semantic camouflage or long-context attacks. In AdvBench and GPTFuzz evaluations covering base and reasoning models and five types of jailbreak attacks, both reduce attack success rate and toxicity.