Claude decodes NIST curve seeds, Apple runs 816B models locally, and Mercor maps AI benchmark gaps

In this issue

Claude identifies Unix timestamps as origin of NIST ECDSA curve seeds
Claude Code team shifts to automated-first code review
PostgreSQL 19 delayed to November after AI code reviews surface more bugs than developers can fix
Mercor finds 90% of US working hours lack AI benchmark coverage
Calif details local root exploit in OpenBSD via 37-year-old integer overflow
Apple demonstrates local inference of near-trillion-parameter models on Mac Studio clusters
and 4 more in brief

Top stories

Claude identifies Unix timestamps as origin of NIST ECDSA curve seeds

Anthropic reports that Claude resolved the decades-old mystery surrounding the origins of NIST elliptic curve seeds. The model discovered that the seeds for P-192, P-224, P-256, P-384, and P-521 were generated using the C library's lrand48 random-number generator, initialized with a Unix timestamp. This finding contradicts previous speculation that the seeds resulted from hashing humorous passphrases. Anthropic states this is not a cryptographic weakness but increases confidence that the curves were chosen without intentional backdoors. The company notes that no action is required for users, as the curves remain secure.

anthropic.com HN discussion

Claude Code team shifts to automated-first code review

The Claude Code team has restructured its code review process to handle the high volume of agent-generated code. They now optimize for the minimum human review time required to prevent reverts and incidents. This approach relies on continuous automated checks and multiple layers of agent-based scrutiny before human approval. The goal is to reserve human attention for complex issues while maintaining code quality and reducing the time spent on routine fixes.

claude.dev HN discussion

PostgreSQL 19 delayed to November after AI code reviews surface more bugs than developers can fix

The PostgreSQL Global Development Group has postponed the release of PostgreSQL 19 to around November 19, reverting eight to ten significant features. Peter Eisentraut, a leading committer, attributes the delay to LLM-assisted code reviews uncovering a long tail of edge-case defects that would previously have been found years later. He notes that while some reverted features had architectural flaws, the AI tools raised the quality bar after code was committed. Eisentraut clarifies that the issue is not "vibe coding," as that technique was unavailable when most features were written, and expects the situation to normalize as AI review capabilities improve.

theregister.com HN discussion

Mercor finds 90% of US working hours lack AI benchmark coverage

Mercor’s US Benchmark Map indicates that 90% of American working hours involve tasks with little or no coverage from existing AI benchmarks. The analysis maps 29 agentic benchmarks against 923 O*NET occupations, finding that only 208 occupations have tasks represented by at least five benchmark prompts. Coverage is not random; it heavily favors high-wage, AI-exposed roles like lawyers while underrepresenting similar lower-wage positions. This gap makes it difficult to assess how well AI models perform in the actual work Americans do.

mercor.com HN discussion

Calif details local root exploit in OpenBSD via 37-year-old integer overflow

Calif published a technical analysis of a local privilege escalation vulnerability in OpenBSD, tracing the root cause to a signed 16-bit shared memory attachment counter defined in original BSD sources. The firm demonstrated that repeatedly forking processes with shared memory attachments could wrap this counter to a negative value, creating an integer overflow. Combined with a race condition and a separate out-of-bounds kernel write, this allowed attackers to achieve root access even with KARL, SMAP, and SMEP enabled. The vulnerabilities were addressed in OpenBSD errata 026 and 018. Calif also proposed defensive improvements, including safe-unlinking implementations and base address randomization for kernel memory allocators to mitigate similar exploitation paths.

calif.io HN discussion

Apple demonstrates local inference of near-trillion-parameter models on Mac Studio clusters

Apple ML Research reports running Muse Spark Open, a model with 816 billion parameters, locally on Mac Studio hardware using the MLX framework. The team benchmarks performance on M5 Ultra and M3 Ultra systems, showing that the new GPU Neural Accelerators deliver up to 4.6 times faster time-to-first-token compared to the previous generation. By leveraging RDMA over Thunderbolt for distributed inference across multiple Mac Studios, the setup achieves over 36 tokens per second, exceeding average human reading speed. This approach allows researchers to experiment with frontier-class models in a private environment without relying on external cloud infrastructure.

machinelearning.apple.com HN discussion

In brief

AVIF outperforms JPEG XL for web images. The author finds that AVIF consistently produces smaller files than JPEG XL for lossy photos, screenshots, and images with transparency. Benchmarks show AVIF also encodes and decodes faster, making it the more efficient choice for most web use cases.

jakearchibald.com HN discussion

C++26 standardizes a debugging header to replace platform-specific breakpoint code. C++26 introduces a new <debugging> header via P2546R5 and P2810R4. This standardizes programmatic breakpoints and debugger detection, replacing manual implementations found in frameworks like Boost.Test and Catch2.

sandordargo.com HN discussion

PS5LM runs llama.cpp on a jailbroken PS5. PS5LM runs llama.cpp on a jailbroken PS5, serving GGUF models via OpenAI and Anthropic APIs. It uses the console's RDNA 2 GPU and supports decision models like d1 Omni 600M.

github.com HN discussion

NIST elliptic curve seeds traced to drand48. A team including researchers from Anthropic claimed the NIST elliptic curve seed bounty. They demonstrated that the seeds were generated using the drand48 random number generator seeded with UNIX timestamps from 1997–98, rather than hashed English sentences.

words.filippo.io HN discussion