SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

DeepMind's AlphaProof Code formally verifies Signal's cryptographic core in four weeks

0
·3 views

Google DeepMind announced AlphaProof Code on November 10, 2026, an agentic framework that formally verified the cryptographic core of the Signal protocol, libsignal v0.105.4 compiled to AArch64 machine code, in four weeks. Agents powered by Gemini 4 Carbon produced over 6.5 million lines of Lean proof, each checked by Lean's kernel. An independent count the following day reported about 7.5 million lines across more than 11,000 files, with no sorry placeholders in the headline proofs. The verification covers one-to-one sessions, the PQXDH handshake, and the Triple Ratchet, but not group chats, linked devices, the operating system, or the underlying cryptographic primitives, which are treated as assumptions. The public repository has been cloned and examined by outside developers, and DeepMind also showed the system catching two deliberately reintroduced bugs.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Seccomp Filters Should Validate Architecture Before Checking Syscall Numbers

A seccomp policy can appear complete yet leave an alternate syscall path open, because filters see only a syscall number and architecture identifier rather than semantic function names. The same integer can mean different things under different ABIs, so a policy checking only the syscall number can be bypassed. Architecture validation should therefore occur at the start of the filter, as shown in Linux seccomp(2) documentation, with unknown architectures rejected. The article recommends choosing default actions matching the risk level, maintaining separate filters and tests for each ABI, and failing closed when instructions cannot be classified. Seccomp is also limited because it cannot inspect pointer contents, so it is usually paired with namespaces, file permissions, and capabilities.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Perfect Annotation Agreement Can Hide Sparse Labelling Coverage

A single 100% agreement figure can make two annotation files appear more complete than they are. In the bundled partial example, only one of twelve records is labelled by both sides, so conditional agreement reads 100% while both-labelled coverage is just 1/12. The metric counts only records labelled by both sides, so it ignores the remaining records entirely. The article's review tool therefore checks that both checkpoints share the same records, taxonomy, and order before comparing them, and it keeps manual decisions separate from the original annotations. It also stresses that agreement is not accuracy and that the two annotation roles are not verified as independent people.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Guide Outlines C# Techniques for High-Throughput RAG Pipelines and Vector Database Tuning

A DEV Community article presents an architectural blueprint for building high-throughput agentic and retrieval-augmented generation pipelines in C#. It covers vector database tuning, including IVF cluster parameters such as nlist and nprobe, alongside HNSW indexing. The piece explains how Span, ReadOnlySpan, and pooled memory buffers can reduce garbage collection overhead during heavy vector operations. It also includes a SIMD-accelerated cosine similarity example and describes a two-tier L1/L2 caching model to limit redundant similarity queries. The author reports benchmark results indicating a reported throughput gain of about 4.8x over default settings when HNSW tuning and caching are combined.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Builds Neural Network Basics in Plain JavaScript to Understand How AI Learns

A JavaScript developer set out to understand how AI models learn after realising that using AI tools was not enough to explain their inner workings. Rather than relying on frameworks like TensorFlow or PyTorch, they chose to implement the fundamentals from scratch in plain JavaScript. They began with a simple task of teaching a program to predict the sum of two numbers from examples, illustrating the core idea of supervised learning. The article then explains how a neuron combines weighted inputs with a bias, and how a loss function measures prediction error so the model can adjust itself. The author presents the approach as a step toward eventually building a GPT-style model.

0 comments Read more at DEV Community