AI Briefing
Hacker News news
The latest AI and developer news from Hacker News, with short summaries.
Visit website →Stream Vessel implements per-user MCP redirect allowlists to prevent OAuth phishing
The design validates redirect URIs at the authorization step rather than registration, allowing users to trust specific https addresses while blocking open-redirect attacks.
Tianhua Chen publishes 195-page mathematical primer on generative AI foundations
The arXiv paper (2605.29713) offers a derivation-oriented guide connecting PCA, VAEs, diffusion models, and GANs for researchers and students.
Kairos: Open-source fraud decisioning platform with agent-first MCP architecture
Apache-2.0 licensed platform that allows coding agents to tune risk models via MCP and supports decision models that operate with few labels.
AmpUp launches Enterprise Brain preview for versioned agent memory
The private preview tool integrates with Claude Code, Codex, and Cursor, claiming to raise root-cause accuracy on ORCA-bench from 6.3% to 35.2% through learned, versioned context.
US Tech Giants' Off-Balance-Sheet AI Debt Reaches $3.2 Trillion
Off-balance-sheet AI-related obligations for major US tech companies nearly doubled in six months to $3.2 trillion, reaching 80% of their assets.
Octomind: Open-source Rust AI agent harness outperforms Claude Code on benchmarks
The Rust-based CLI tool achieves a 24/25 task success rate at $63.43, beating Claude Code's 23/25 at $81.79 in a July 2026 benchmark.
New scaling law prescribes quadrillion-parameter models for data-constrained LLM training
Baron et al. propose a scaling law that recommends ~7,000x more parameters than Chinchilla at 10^29 FLOPs, using hyperparameter transfer and ensembling to maintain loss reduction on fixed datasets.
Anthropic S-1 filing reveals $1.7B quarterly personnel costs and $107K average monthly pay
Anthropic's S-1 filing reveals $1.7B in Q3 2026 personnel costs, with stock-based compensation representing the largest share of the expense.
NVIDIA releases Auto Ontology, an open-source framework for agent data context
The Apache 2.0 licensed tool generates ontologies from enterprise metadata to provide trusted context for AI agents via MCP, CLI, and API interfaces.
Replay QA uses time-travel debugging to identify root causes of web app bugs
Replay QA, built on a deterministic time-travel debugger for Chromium and Node.js, identifies over 900 bugs in 40 apps by providing agents with precise execution evidence rather than surface-level logs.
Revolut shifts 2,000 engineers to self-hosted GLM, cutting AI costs to a tenth of previous spend
The bank deployed 56 Nvidia B200s to run GLM internally, reducing token costs to 30 cents per million and improving engineering lead time by 70%.
Researchers Demonstrate Remote Code Execution Backdoors in LLM Agents via Pretraining Poisoning
Poisoning just 0.1% of the pretraining corpus for Qwen3 models implants a backdoor that allows agents to execute attacker-chosen commands on user machines.
Momentic Launches AI QA Agents for End-to-End App Testing
The YC W24 startup uses AI agents to detect bugs in critical app flows, reporting 117,010 detected bugs and 70.6 million test runs across its user base.
Goldman Sachs Restricts Claude Code and GitHub to Engineers
The bank's October internal guidance limits third-party AI coding agents to its engineering staff, excluding traders and bankers who previously used them for client flow monitoring and market modeling.
PickLongCat-2.5 Launches with Native Multimodality and 1.6T Parameters for Agentic Coding
The 1.6T parameter model achieves 88.4 on Terminal Bench v2.1 and 91.9 on BrowseComp, outperforming LongCat-2.0 and approaching frontier models like Claude-Opus-5.
OpenAI releases GPT-Rosalind-6 for trusted biological research access
The model shares GPT-6 Astra weights and achieves 49.6% on MedChemBench v7, exceeding the High capability threshold in three of four biological evaluations while remaining below Critical thresholds.
PickAnthropic amended S-1 reveals $17 billion Q3 revenue and $895.4 billion in compute commitments
The filing reports $33.24 billion in nine-month revenue, a 1,200% year-over-year increase, and a $104 billion net loss driven largely by fair value changes.
MIT CSAIL Research Demonstrates LLMs Can Natively Read and Write Neural Network Weights
MIT CSAIL researchers demonstrate that GPT-6 Astra can read and write neural network weights, improving Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by editing 0.016% of its parameters to fix reasoning loops.
bf16 Fine-Tuning Freezes 95% of Weights, Masking Optimizer Instability
Using bf16 weights with AdamW without fp32 master copies causes large-magnitude weights to freeze, effectively acting as a sparse fine-tuner that masks underlying training instability.
Anthropic S-1A filing warns of imminent recursive self-improvement threshold
The filing states internal estimates suggest future models may soon meet the RSI definition, while current models remain below the threshold with zero fully autonomous research share.