Research

Uniform activation steering is a blunt instrument that wrecks LLM utility on benign inputs. Apple's DSAS fixes that — with complexity strings attached.

Proofcraft completes security proofs for seL4 on AArch64, providing formal mathematical proof of confidentiality, integrity, and functional correctness. This milestone ensures the kernel prevents unauthorized access and data modification, guaranteeing security isolation for applications.

Formal verification is a tool to verify correctness of mathematical statements. In this tutorial, we'll write simple statements about cryptography and their proofs in Lean, focusing on the One-Time Pad protocol.

Exploit brokers pay $500k for WordPress RCEs. I found one with GPT5.6 and $25. The vulnerability was discovered using a large language model and can be exploited for remote code execution.

Moonshot AI releases Kimi K3, a 2.8-trillion-parameter open model with native multimodality and a Mixture-of-Experts architecture. The model achieves approximately 2.5× better scaling efficiency than its predecessor and is available via Kimi.com, Kimi Work, Kimi Code, and the Kimi API.

Researchers at Anthropic have discovered a 'global workspace' in language models, which could have implications for AI safety and interpretability. The workspace, dubbed the 'J-space', is a small collection of internal neural patterns that play a special role in the model's processing.

SQLite recently patched a rare, 16-year-old bug in its Write-Ahead Log (WAL) checkpointing system that could lead to database corruption. This post from Canonical's dqlite (distributed SQLite) team walks through how they used TLA+ to formally model SQLite's internal behavior, isolate the exact sequence of steps needed to trigger the corruption, and determine whether their own system was vulnerable.

Mistral’s Leanstral 1.5 delivers a 119 B‑parameter mixture‑of‑experts tuned for Lean 4 theorem proving, with a 256 k token window and zero per‑token cost in Labs.