arXiv:2608. 07534v1 Announce Type: cross Abstract: Ultraconstructive Model Theory (UCMT) replaces idealized satisfaction, at finite compu- tational scale, by bounded adversarial survival.
arXiv:2611. 06303v1 Announce Type: cross Abstract: The halting problem asks whether a program eventually halts, with unrestricted resources.
arXiv:2606. 13621v1 Announce Type: new Abstract: Shielded reinforcement learning is typically presented as a runtime safety mechanism that compiles temporal-logic specifications into automata restricting an agent's actions.
arXiv:2605. 07032v2 Announce Type: replace-cross Abstract: The evolution of generative models from next-token predictors to autonomous engines of complex systems necessitates rigorous safety hardening.
The paper examines the vulnerability of open‑weight large language model safeguards that aim to block harmful usage by assuming new harmful behavior arises only through fine‑tuning. It demonstrates that simple, low‑cost attacks—abliteration and prefilling—can significantly increase attack success rates across multiple benchmarks, and proposes an abliteration‑resistant tuning (ART) method that mitigates these attacks. The study shows that safeguarding defenses must be evaluated against a broader range of attack strategies beyond fine‑tuning.
The paper investigates whether increased computational resources for an attacker can make adversarial attacks transferable across all defenders. It finds that for classification tasks, a quadratic resource gap guarantees either a transferable attack or a defense, whereas for generative tasks this guarantee fails. The authors also show that a defense that adds computation after seeing the input can be provably stronger than any fixed-model defense, suggesting that inference-time compute can enhance robustness.
Shielded reinforcement learning is typically presented as a runtime safety mechanism that compiles temporal-logic specifications into automata restricting an agent's actions. We argue this is the wrong product.
arXiv:2606. 30107v1 Announce Type: new Abstract: An unreliable language model can be made to produce reliable physical designs if the authority to assert is moved out of the model: the model proposes, and a deterministic engine alone certifies, returning certified, impossible, or unknown.
arXiv:2606. 28639v2 Announce Type: replace-cross Abstract: We establish the mathematical limits of AGI safety in two forms: verifying a fixed system, and verifying that a certified safety property persists once the system self-modifies.
arXiv:2610.28045v1 Announce Type: new Abstract: To defend against backdoor attacks on neural networks, the defender must identify the verifier function that determines whether an input is a trigger,...
AnchorRep is a lightweight LoRA adapter that defends large language models (LLMs) against cross-model adversarial transfer by repelling internal representations of harmful prompts away from those of a frozen anchor model. It requires only a small set of harmful prompts for training and achieves a cross‑model attack success rate of 1.1% across five models and four architectural families. Compared to existing defenses, AnchorRep avoids high costs such as up to 77% degenerate benign output or 18% over‑refusal, and introduces the Benign Garble Rate metric to capture these failures.
arXiv:2608. 11815v1 Announce Type: new Abstract: Transfer-based adversarial attacks craft adversarial examples using surrogate models to mislead black-box victim models.
arXiv:2602. 18396v2 Announce Type: replace Abstract: We propose PRISM-FCP (Partial shaRing and robust calIbration with Statistical Margins for Federated Conformal Prediction), a communication-efficient Byzantine-robust federated conformal prediction framework that uses partial model sharing to mitigate stochastic model-poisoning attacks during training and histogram-based filtering to mitigate adversarial calibration submissions.