Skip to content
devup↗
Site language
TürkçeEnglish

Edition #4 · 2026-10-17

NIST curve seeds solved, OpenBSD kernel bug and MLX inference

The decades-old mystery behind NIST elliptic curve seeds is resolved alongside a 37-year-old OpenBSD kernel flaw and new local LLM tooling.

  1. SecurityHacker News

    Anthropic researchers resolve origins of NIST elliptic curve seeds

    Anthropic researchers have solved the decades-old mystery behind the NIST P-192, P-224, P-256, P-384, and P-521 elliptic curve seeds specified in FIPS 186-2. Long suspected of potentially concealing NSA backdoors, the seeds were revealed not to be hashes of secret passphrases, but outputs of the standard C library random number generator (lrand48/drand48) seeded with Unix timestamps from 1997. The discovery does not weaken ECDSA or introduce new cryptographic vulnerabilities; instead, it reinforces confidence that the curves lack intentional subversion.

    Read original
  2. SecurityHacker News

    Calif discloses 37-year-old privilege escalation bug in OpenBSD kernel

    Security researchers at Calif documented a local privilege escalation flaw in OpenBSD tracing back roughly 37 years to early BSD sources. The SysV shared memory attachment counter was originally defined as a signed 16-bit integer, allowing repeated process forking to push the counter past 32,767 and wrap it into negative numbers. This integer overflow causes a use-after-free race condition that, when chained with a separate out-of-bounds kernel write, achieves local root execution despite KARL, SMAP, and SMEP exploit mitigations.

    Read original
  3. AIHacker News

    Apple ML Research demonstrates frontier LLM inference with MLX on Mac Studio

    Apple Machine Learning Research published benchmarks showing local and distributed LLM inference on M5 Ultra Mac Studio hardware using the MLX framework. Evaluating the 816-billion-parameter Muse Spark Open mixture-of-experts model in NVFP4 precision across a cluster of four Mac Studios linked via RDMA over Thunderbolt, the setup generated over 36 tokens per second at 8,192 context. On a single 512GB Mac Studio, models like Mistral Medium 3.5 (128B) and Qwen 3.8 (27B) ran natively in full FP8 and BF16 up to 65,536 context sizes.

    Read original
  4. SoftwareHacker News

    Claude Code team shifts code review to automated multi-layer agent checks

    The Claude Code team outlined its operational shift away from manual, line-by-line human reviews to manage high volumes of agent-generated code. From June to September, merged pull requests doubled while revert rates dropped by over 40 percent. Rather than relying on peer review gates at PR creation, the team runs continuous pre-PR verification where subagents review diffs, execute tests, and evaluate UI screenshots, reserving human developer intervention primarily for architectural decisions and edge cases where rollback guardrails apply.

    Read original
  5. Open sourceLobsters

    Peer-to-peer networking library Iroh reaches version 1.0

    After four years of development, peer-to-peer networking project Iroh released version 1.0. Originally started as a Rust-based IPFS implementation targeting resource-constrained devices, the team pivoted away from IPFS compatibility to eliminate abstraction overhead and heavy dependency trees. The final 1.0 architecture standardizes on endpoints identified by Ed25519 public keys, BLAKE3 verified streaming for blob transfers, and Quinn-based QUIC as its sole transport layer, offering an embeddable, low-overhead sync engine.

    Read original
  6. AIHacker News

    Black Forest Labs releases open-weights FLUX 3 Image Klein

    Black Forest Labs released FLUX 3 Image Klein, an open-weights 5.6-billion-parameter image generation and editing model under the FLUX Kommunity License. Supporting native 4K output, bounding-box placement controls, and up to 10 reference images, the model delivers top ratings against larger open baselines. Optimized alongside NVIDIA, its INT8 and distilled configurations run locally in ComfyUI on consumer GeForce RTX GPUs with as little as 8GB of VRAM, generating 1MP images in roughly 4 seconds on an RTX 5090.

    Read original
  7. Open sourceLobsters

    Linus Torvalds rejects proposed bans on AI-assisted kernel contributions

    Responding on the Linux kernel mailing list, Linus Torvalds firmly rejected blanket bans on AI-assisted patches, arguing that restrictions on tooling are nonsensical and that maintainers should judge contributions strictly by technical comprehension and code quality.

    Read original
  8. SoftwareLobsters

    Rust 1.99.1 point release fixes FFI miscompilation and linker regressions

    The Rust team issued point release 1.99.1, fixing a missing wcslen link regression on UEFI std targets, restoring required zero-extension ABI handling for x86_64 FFI bool returns, and adding an rpath entry for macOS rust-lld execution.

    Read original
  9. ToolsHacker News

    PS5LM runs local GGUF models on jailbroken PlayStation 5 consoles

    The open-source PS5LM project enables jailbroken PlayStation 5 consoles to host GGUF language and decision models locally via llama.cpp's Vulkan backend on the console's RDNA 2 GPU, exposing OpenAI and Anthropic-compatible local endpoints.

    Read original
  10. SoftwareHacker News

    C++26 standardizes programmatic breakpoints with <debugging> header

    Standard proposals P2546R5 and P2810R4 introduce the <debugging> header to C++26, standardizing cross-platform debugging primitives with std::breakpoint(), replaceable std::is_debugger_present(), and std::breakpoint_if_debugging().

    Read original
All editions →