arXiv:2607. 26639v1 Announce Type: cross Abstract: A self-check defense asks the target model to assess a request before answering it; SAGE, the strongest published instance, reports an average 99% defense success rate.
A self-check defense asks the target model to assess a request before answering it; SAGE, the strongest published instance, reports an average 99% defense success rate. We show it can be breached by composing two attacks that are individually harmless against it: an established code-completion encoding and an established best-of-N search, neither of which exceeds 4.
The paper investigates how allocating a query budget to structural depth rather than surface variation improves jailbreak success against the SAGE self‑check defense. By using a best‑of‑N approach over a code‑completion encoding, the authors achieve 67%, 22%, and 15% success rates on three open‑weight targets—far exceeding the 4.7% and 3.0% rates of single‑draw encoding and character‑search methods. The study demonstrates that depth of encoding and breadth of variation independently undermine transform and gate defenses, and that repeated sampling can inflate perceived robustness.
arXiv:2607. 26574v2 Announce Type: replace-cross Abstract: Safety classifiers ("guards") are the dominant black-box defense for vision-language models, yet a guard judges an input's surface form, not its meaning: a harmful request re-encoded as set theory, formal logic, a classical language, code, or text rendered inside an image slips past a guard that would block it in plain language - the decode gap.
arXiv:2607. 26574v1 Announce Type: cross Abstract: Safety classifiers ("guards") are the dominant black-box defense for vision-language models, yet they judge an input's surface form, not its meaning: a harmful request re-encoded as set theory, formal logic, a rare language, code, or an image of text slips past a guard that would block it in plain language -- the decode gap.
arXiv:2603. 24511v2 Announce Type: replace-cross Abstract: We show that AI agents are capable of discovering novel algorithms for adversarial attacks against LLMs, advancing the state of the art on white-box jailbreaking and prompt injection evaluations.
The paper critiques current encoded‑prompt safety benchmarks that focus only on harmful requests, showing that such tests can misrepresent a model’s safety. By evaluating the benign arm under the same encoding, the authors reveal a substantial drop in the harm gap—sometimes to zero—indicating that the encoding masks true safety deficiencies. Across multiple large models and training pipelines, they document that the encoding can either hide or falsely inflate safety metrics, and they identify twelve specific instrument defects that contribute to these misleading results.
arXiv:2606. 29441v1 Announce Type: cross Abstract: Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists.
The paper introduces RedHerring, a defense mechanism that inserts safe decoy vulnerabilities into code repositories to divert autonomous LLM agents’ verification efforts away from real security flaws. By embedding CVE-derived vulnerability chains with false bridges and providing a private certificate for quick verification, RedHerring forces agents to spend a significant portion of their limited resources on decoys. Experiments on 33 OSS‑Fuzz projects show a 38.7‑60.4% reduction in discovered real vulnerabilities, even when agents are aware of decoys.
arXiv:2605. 03226v2 Announce Type: replace-cross Abstract: Safety fine-tuning of language models typically requires a curated adversarial dataset.
The paper presents a preprocessor that recovers and decodes encoded content in vision‑language models to close the decode gap that allows harmful requests to bypass safety classifiers. Evaluated against eleven encoding attacks, the preprocessor raises block rates from 0 % to 67‑90 % but also increases benign over‑refusal, and no configuration achieves an ensemble attack‑success rate below 40 % while keeping benign over‑refusal under 70 %. The study shows that closing one encoding channel merely relocates success rather than eliminating it, highlighting the limits of recovery‑based defenses.
The paper introduces ‘Fool’s Gold’, a defensive deception technique for open‑weight language models that hardens them against safety‑removal attacks. By training decoy responses within a differentiable simulation of the attack, the method poisons the payoff of stripped refusal mechanisms, producing confident but falsified answers to hazardous requests while preserving benign behavior. Experiments on seven models (9B‑122B) show that 51‑90% of attacked‑state responses become decoys, with the defense accounting for 27‑84% of this effect, and that the defended 122B model remains within benign‑behavior budgets.
The paper investigates whether the hidden activations of large language models (LLMs) contain signals about the vulnerability of C/C++ code when the code is provided as context. By extracting prefill token activations from four LLMs and training small MLP probes, the authors achieve an average F1 score of 41.7% across four benchmarks, with the best probe matching state‑of‑the‑art fine‑tuned classifiers on the Devign dataset. The results suggest that a coding LLM’s internal representation can inform vulnerability detection, opening the door to lightweight, model‑native screening methods.