Trojan-Speak: Bypassing Constitutional Classifiers with No Jailbreak Tax via Adversarial Finetuning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Sel, Bilgehan, He, Xuanli, Peng, Alwin, Jin, Ming, Wei, Jerry
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917372060762112
author Sel, Bilgehan
He, Xuanli
Peng, Alwin
Jin, Ming
Wei, Jerry
author_facet Sel, Bilgehan
He, Xuanli
Peng, Alwin
Jin, Ming
Wei, Jerry
contents Fine-tuning APIs offered by major AI providers create new attack surfaces where adversaries can bypass safety measures through targeted fine-tuning. We introduce Trojan-Speak, an adversarial fine-tuning method that bypasses Anthropic's Constitutional Classifiers. Our approach uses curriculum learning combined with GRPO-based hybrid reinforcement learning to teach models a communication protocol that evades LLM-based content classification. Crucially, while prior adversarial fine-tuning approaches report more than 25% capability degradation on reasoning benchmarks, Trojan-Speak incurs less than 5% degradation while achieving 99+% classifier evasion for models with 14B+ parameters. We demonstrate that fine-tuned models can provide detailed responses to expert-level CBRN (Chemical, Biological, Radiological, and Nuclear) queries from Anthropic's Constitutional Classifiers bug-bounty program. Our findings reveal that LLM-based content classifiers alone are insufficient for preventing dangerous information disclosure when adversaries have fine-tuning access, and we show that activation-level probes can substantially improve robustness to such attacks.
format Preprint
id arxiv_https___arxiv_org_abs_2603_29038
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Trojan-Speak: Bypassing Constitutional Classifiers with No Jailbreak Tax via Adversarial Finetuning
Sel, Bilgehan
He, Xuanli
Peng, Alwin
Jin, Ming
Wei, Jerry
Cryptography and Security
Artificial Intelligence
Computation and Language
Fine-tuning APIs offered by major AI providers create new attack surfaces where adversaries can bypass safety measures through targeted fine-tuning. We introduce Trojan-Speak, an adversarial fine-tuning method that bypasses Anthropic's Constitutional Classifiers. Our approach uses curriculum learning combined with GRPO-based hybrid reinforcement learning to teach models a communication protocol that evades LLM-based content classification. Crucially, while prior adversarial fine-tuning approaches report more than 25% capability degradation on reasoning benchmarks, Trojan-Speak incurs less than 5% degradation while achieving 99+% classifier evasion for models with 14B+ parameters. We demonstrate that fine-tuned models can provide detailed responses to expert-level CBRN (Chemical, Biological, Radiological, and Nuclear) queries from Anthropic's Constitutional Classifiers bug-bounty program. Our findings reveal that LLM-based content classifiers alone are insufficient for preventing dangerous information disclosure when adversaries have fine-tuning access, and we show that activation-level probes can substantially improve robustness to such attacks.
title Trojan-Speak: Bypassing Constitutional Classifiers with No Jailbreak Tax via Adversarial Finetuning
topic Cryptography and Security
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2603.29038