1st Workshop on Hidden Information, Steganography, Privacy, and Emerging Risks (WHISPER 2026), co-located with the 2nd Joint
National Conference on Cybersecurity (ITASEC & SERICS 2026)
Workshop
Large Language Models (LLMs) are nowadays a core technological enabler for several documental workflows, for
instance, for preparing summaries or correcting errors both in code and prose. Apart from efficiency gains, their
utilization also introduces a wide range of security hazards. For the specific case of scientific pipelines, LLMs may
be exploited to disrupt the review process, thus requiring suitable workarounds. Therefore, this paper investigates
the robustness of a commercial LLM (specifically, OpenAI GPT-5.1) in different scenarios where an attacker injects
a prompt to alter the outcome of the review. Our work analyzes the effectiveness of various encoding strategies
for bypassing detection and filtering mechanisms, as well as prospective countermeasures. Preliminary results
confirm that GPT-5.1 is permeable to crafted payloads but also capable of recognizing many attempts. Moreover,
our trials indicate that preparing a suitable prompt for “sanitization” can be a prime mitigation strategy.
Read more































