Research map / Adversarial robustness and security / Attacks on text and graphs
Trojan-Speak: Bypassing Constitutional Classifiers with No Jailbreak Tax via Adversarial Finetuning
Bilgehan Sel, Xuanli He, Alwin Peng, Ming Jin, Jerry Wei
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add hybrid RL+SFT training after SFT-only stages (ablation) | Trojan-Speak on Haiku 4.5 · Bug Bounty Benchmark | rubric score Avg@5 (%): SFT-only 55-65 -> 85.2 | Improved |
| Interleave SFT steps with RL instead of pure RL (ablation) | Haiku 4.5 RL on cipher-encoded questions · Nemotron-MCQA encoded accuracy | accuracy | Improved |
| Use minimal frequency-ordered substitution cipher instead of Base64 or ROT13 (ablation) | Haiku 4.5 · GPQA-Diamond | accuracy (%): 41.4 (Base64) -> 62.6 | Improved |
| Use Base64 with case swap rather than minimal cipher (ablation) | Haiku 4.5 · GPQA-Diamond | accuracy (%): 62.6 -> 35.7 | Hurt |
| Substitute frequency-ordered letters (7+) instead of random letters (ablation) | substitution cipher · CBRN query classifier bypass | bypass rate (%): consistent 99+% with 7+ substitutions | Improved |
| Use fewer substituted letters (ablation) | cipher complexity · GPQA-Diamond (Qwen3 14B, 32B) | capability retention | Improved |
| Vary template structure and chunk length (32 to 256) (ablation) | Trojan-Speak on Qwen3 14B · GPQA-Diamond | accuracy: standard deviation 0.53%, max spread 1.5% | No clear effect |
| Increase LoRA rank (8 to 128) (ablation) | Trojan-Speak on Qwen3 8B/14B/32B · GPQA-Diamond during Stage 2 | accuracy | Improved |
| Use smaller base model (Qwen3 8B/4B vs 14B+) (ablation) | Trojan-Speak · GPQA-Diamond | capability retention (%): 89% (14B) -> 53% (8B), 55% (4B) | Hurt |
| Apply Trojan-Speak fine-tuning to Haiku 4.5 | Claude Haiku 4.5 · GPQA-Diamond | accuracy (%): 65.7 -> 62.6 | Hurt |
| Train activation probe on cipher-trained model rather than base model (ablation) | Softmax and SWiM probes · Ciphered harmful content detection | AUC: 0.27-0.46 -> 0.97-0.99 | Improved |
| Add probe-evasion training strategies to the attack (ablation) | Trojan-Speak · Activation probe detection | probe detection performance | No clear effect |
Closest papers by meaning
- Fundamental Limitations in Pointwise Defences of LLM Finetuning APIs (NeurIPS 2025 · similarity 0.85)
- Metis: Learning to Jailbreak LLMs via Self-Evolving Metacognitive Policy Optimization (ICML 2026 · similarity 0.85)
- Covert Malicious Finetuning: Challenges in Safeguarding LLM Adaptation (ICML 2024 · similarity 0.84)
- Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs (NeurIPS 2026 · similarity 0.84)
- Breach By A Thousand Leaks: Unsafe Information Leakage in 'Safe' AI Responses (ICLR 2025 · similarity 0.84)
- SafeHarbor: Defining Precise Decision Boundaries via Hierarchical Memory-Augmented Guardrail for LLM Agent Safety (ICML 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
