The generation of high-fidelity, controllable malicious network traffic is essential for simulating realistic cyberattacks, evaluating defense mechanisms, and enhancing intrusion detection systems (IDS). However, existing approaches often suffer from low protocol fidelity, limited structural control, and poor adaptability, reducing their effectiveness in practical cybersecurity applications. We introduce TrafficT5, a three-stage, self-correcting framework that turns natural-language intents into executable PCAPs. It (i) predicts flow-level features, (ii) generates byte-aligned hex under a fixed 00–FF vocabulary, and (iii) invokes a repair module that deterministically enforces protocol invariants and performs detector-guided, iterative byte-level correction trained with multi-task objectives. The result is traffic that is both semantically coherent and protocol-compliant. Extensive evaluations on five network datasets demonstrate that TrafficT5 achieves an average Bad Packet Rate (BPR) of only 0.32%, significantly outperforming existing methods. Furthermore, synthetic malicious flows generated by TrafficT5 are reliably detected as threats by mainstream IDSs, and augmenting datasets with these flows improves F1 detection scores by over 5% under low-resource conditions.

While Tor provides strong anonymity, it also facilitates the concealment of malicious activities, which poses a significant challenge to cybersecurity surveillance. As an effective anti-anonymity technique, Website Fingerprinting(WF) enables the inference of which websites a user is visiting, thereby uncovering potential attacker activities. State-of-the-art(SOTA) methods have demonstrated remarkable effectiveness. However, a large number of labeled traffic is required to ensure effectiveness, and without timely updates, these models will encounter serious challenges of concept drift due to the dynamic nature of website content and network conditions. The core reasons lie in the independently and identically distributed assumption, while in challenging open-world scenarios, the long-term spatial and temporal dynamics complicates data consistency and effective knowledge transfer. To address these issues, this paper presents WF-TFC, an open-world few-shot anonymous WF model via self-supervised contrastive learning and time-frequency consistency. It aligns time- and frequency-based representations in the latent time-frequency space, enhancing the sustained effectiveness of inherent patterns across various websites. Consequently, it accommodates diverse few-shot target domains with varying dynamics, facilitating data consistency and knowledge transfer in unobserved long-term temporal and spatial environments. For instance, with only 5 traces per website, WF-TFC achieves 92.62% accuracy on traces collected six weeks after pre-training, exceeding the SOTA(i.e., NetCLR) by 2.12%. On similar but mutually exclusive traces, it attains an F1 score of 87.20%, surpassing the SOTA by 6.12%.