AI
Controlled Decoding Attacks on Black-Box LLMs
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 01:19:53 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 04:18:25 UTC.
Briefings
A new framework called Sample-Based Distribution Reconstruction enables jailbreaking via text-only interfaces by reconstructing model probabilities from sampled outputs. This approach avoids the high query costs of repeated sampling by concentrating control on a small subset of positions. The method uses risk-gated residual control to decide when to reconstruct and speculative multi-token execution to amortize target calls. Across four endpoints and three benchmarks, the system achieved the highest mean score among all comparisons against baselines.
Key points
- The framework Sample-Based Distribution Reconstruction reconstructs model probabilities from sampled outputs to bypass safety alignment without access to weights.
- Risk-Gated Residual Control uses the evolving response prefix to decide when to reconstruct and modify the distribution, concentrating sampling costs at selected positions.
Context and limitations
The approach requires repeated sampling and assistant-prefix continuation, which incurs substantial query costs. It is limited to interfaces returning only sampled text, as it does not apply to those requiring access to model weights or numerical token probabilities.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 01:19:53 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeihsuh4wye4hyugg3exwwckpe5ortlguavjml5bxlq457ia7r4u6mu