Research map / Large language models / Agents, code and math
Early Signs of Steganographic Capabilities in Frontier LLMs
Artur Zolkowski, Kei Nishimura-Gasparian, Robert McCarthy, Roland S. Zimmermann, David Lindner
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 |
|---|---|---|---|
| Allow unmonitored CoT for encoder and decoder (ablation) | Frontier LLMs, message passing without coordination · Message passing past GPT-4.1-mini monitor | transmission success | Improved |
| Let encoder and decoder share the encoding scheme, with CoT (ablation) | Reasoning models and GPT-4.5 · Message passing with monitor | monitor bypass rate | Improved |
| Remove CoT in coordinated message passing (ablation) | Frontier LLMs · Message passing with monitor | success | Hurt |
| Use uncommon rather than common encoding schemes (no cover task) (ablation) | Frontier LLMs · Encoding and decoding messages | implementation accuracy | Hurt |
| Increase state-tracking sequence length (first-letter, no cover task) (ablation) | GPT-4.5 · State-tracking 1D | accuracy: 100% -> 90% -> 67% | Hurt |
| Embed blatant model-generated scheme within a cover task (ablation) | GPT-4.5 · State-tracking 1D, 12 moves | accuracy: 97% -> 73% | Hurt |
| Use subtle model-generated encoding prompts (ablation) | GPT-4.5 · State-tracking 1D | accuracy | No clear effect |
| Use common encodings (Base64/Hex) instead of unrestricted reasoning on AQuA-RAT (ablation) | GPT-4.1 · AQuA-RAT | accuracy: 88% -> 47-78% | Hurt |
| Give a specific steganography scheme (strong hint) | GPT-4.1 · Case-study exfiltration and reward hacking | goal success: 90% and 59% | Improved |
Closest papers by meaning
- Large language models can learn and generalize steganographic chain-of-thought under process supervision (NeurIPS 2025 · similarity 0.89)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.85)
- On Evaluating the Durability of Safeguards for Open-Weight LLMs (ICLR 2025 · similarity 0.84)
- Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory (ICLR 2024 · similarity 0.84)
- Bits Leaked per Query: Information-Theoretic Bounds for Adversarial Attacks on LLMs (NeurIPS 2025 · similarity 0.84)
- Unbiased Watermark for Large Language Models (ICLR 2024 · similarity 0.84)
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.
