Jailbreaks keep finding new doors: style, bitstreams, split intent, and sampling
Four preprints posted this week attack safety from four different directions, and each one shows the current defenses watching the wrong place.
- Stylistic rewrites disrupt refusal representations in LLaMA-3.1-8B while preserving harmful intent. Self-reported m:…1451
- Style-aware Direct Preference Optimization cuts the attack success rate from 86% to 19.5% while preserving utility. Self-reported m:…1451
- BitBypass bypasses safety alignment in GPT-4o, Gemini 1.5, Claude 3.5, Llama 3.1, and Mixtral. Self-reported m:…1435
- The Split-Intent Jailbreak Attack reaches a 91.38% attack success rate on Qwen-Image-3.0-Pro and 68.28% on GPT Image 2. Self-reported m:…1452
Rewrite the style, keep the intent
The first new attack surface is style itself. A preprint on jailbreaking finds that linguistic style is a systematic and previously overlooked attack surface for large language model safety. Stylistic rewrites disrupt refusal representations in LLaMA-3.1-8B while preserving harmful intent: the model still understands what is being asked, but the rephrasing switches off the refusal. The paper also tests a defense: style-aware Direct Preference Optimization reduces the attack success rate from 86% to 19.5% while preserving utility. m:…1451
The result says safety training is partly memorizing the shape of a bad request, not the meaning. Change the shape and the guard falls, which is why a defense trained on style variation works where a fixed filter does not. m:…1451
The defense result is as important as the attack. Style-aware Direct Preference Optimization cuts the attack success rate from 86% to 19.5% while preserving utility, which suggests the refusal behavior can be retrained to survive rephrasing rather than bolted on afterward. The gap between those two numbers is the measure of how much safety was style-deep: most of the vulnerability disappears once the model is taught that a request is the same request in different clothes. m:…1451
Hide it in bits, split it across modes
The second attack hides the request in a bitstream. BitBypass, a black-box jailbreak, uses hyphen-separated bitstream camouflage to bypass safety alignment, and the authors report it outperforms several state-of-the-art jailbreak attacks in stealthiness and attack success. It successfully bypasses safety alignment in GPT-4o, Gemini 1.5, Claude 3.5, Llama 3.1, and Mixtral. m:…1435
The third splits the intent across image and text. The Split-Intent Jailbreak Attack achieves an attack success rate of 68.28% on GPT Image 2 and 91.38% on Qwen-Image-3.0-Pro. The proposed defense, editing-intent reconstruction, evaluates the complete transformation implied by the joint multimodal context and substantially improves rejection of split-intent attacks. m:…1452
Both attacks work the same way: they give each safety check a piece that looks harmless on its own. The bitstream looks like data, not a request; the image and the text each look innocent until they are combined. A filter that reads one channel at a time will keep missing them. m:…1435 m:…1452
Sampling is not the law
The fourth paper revises what teams thought they knew about sampling attacks. The power-law scaling of best-of-N jailbreak attack success rates is a finite-size effect of the adversarial dataset, the authors argue, not a law of the attack. A five-parameter barrier model can accurately predict the attack surface across sampling counts and temperatures, covering the (N, M) surface and its temperature dependence. m:…1450
For defenders, the four papers point to the same shift: test the model against rephrased, encoded, and split versions of the same request, not just the request itself, and treat a defense that only works on one phrasing as a draft, not a result. m:…1451 m:…1452
The barrier model matters because it changes what a safety team should forecast. If best-of-N success followed a power law, every extra sample would keep buying the attacker more success at a predictable rate. If it is a finite-size effect of the adversarial dataset, the curve bends, and a five-parameter model that predicts the attack surface across sampling counts and temperatures gives defenders a way to estimate risk at volumes they have not tested yet, instead of extrapolating a law that was never a law. m:…1450
What is still unresolved
- Whether style-aware training holds up against styles the defense never saw
- Whether bitstream and split-intent defenses can be combined without slowing the model
- How far the barrier model for best-of-N attacks predicts beyond the datasets it was fit on
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.