All news

Cloudflare cut over 100 TB of RAM from its load balancer by optimizing its Rust consistent hashing implementation.

Attackers built an autonomous agent that hijacks exposed AI model endpoints and runs up to 100,000-step attack chains.

SARIA keeps agent KV caches across turns with coordinated DRAM and SSD reload to cut prefill latency.

MirrorBtr speeds up B-Tree traversals in MySQL InnoDB by adding a learned index alongside the tree instead of replacing it, so transactions and recovery stay untouched while lookups skip most inner nodes.

Evergreen checks claims in LLM-written summaries of tables by compiling each claim into a semantic verification query with citations. It keeps F1 at 0.94 while cutting cost 3.1x through early stopping, relevance sorting and prompt caching.

PRISM is an integer set structure that splits 64-bit keys into prefix segments in a trie with bitmap nodes, pruning whole key ranges during intersection and mapping leaves onto SIMD registers for fast k-way intersection.

vLLM fits long agentic contexts for GLM-5.3 on 16 H200s by splitting prefill and decode nodes and using prefill and decode context parallelism plus hybrid sparse attention.

The Black-Hole Attack poisons vector databases by inserting a few vectors near the centroid of the stored embeddings. Hubness in high dimensions then pulls them into the top-k for up to 94.4% of queries.

pgEdge digs into Postgres numeric internals to explain why aggregates over numeric are much slower than over integer or float types, and what could make them faster.

vLLM serves DeepSeek-V4.1-Flash on one GB300 NVL72 rack at 954K tokens per second per GPU using prefill and decode disaggregation and wide expert parallelism, 7.9 times its single-node result.

Meta describes the first phase of Prometheus, its gigawatt-scale training supercluster behind the Muse models: power, networking and how it scales across buildings.

Part 2 of how OpenAI scaled its storage platform Habitat to serve over 1 billion ChatGPT users across 40 regions with millisecond latency.

NOMA treats multi-agent LLM pipelines as a query optimization problem, jointly choosing topology, model and execution engine against latency, cost and accuracy instead of hand-wiring fixed agent graphs.

ThinkingBox turns database end-state checks into RL rewards for post-training agents, with training code released.

Proposes SLOs and error budgets to govern how much autonomy production LLM agents get, replacing signals built for deterministic services.

A field report showing that recovery_prefetch on Postgres replicas can quietly cause outages, how it was diagnosed, and when to turn it off.

Meta made data pipeline agents more reliable by keeping the orchestrator thin: it only routes work and enforces checks, while specialist agents hold the knowledge.

Cloudflare reworked range-request caching so a video seek no longer pulls a 128 MiB chunk from origin; it now fetches and caches just the needed ranges for large files.

FrontierCS 2 benchmarks agents on frontier computer science research tasks.

EpiSEM-Bench tests whether tool-using agents notice missing safety evidence before acting.

AgentPhishing benchmarks LLM email agents under adaptive phishing.

New PHP2Shell exploit defeats ASLR on x86-64 servers.

Applying the actor model to production AI agent runtime problems like state and failure.

RemoteANN does nearest-neighbor search directly on object storage with up to 16.1x fewer requests and 5.9x higher throughput than baselines.

Separates task decomposition from typed-state QA to explain gains in multi-agent LLM pipelines.

Tests whether self-evolving agents like ReasoningBank, GEPA and Meta-Harness learn new tasks without forgetting old ones.

PHP shipped 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix two critical unauthenticated remote code execution bugs; Amazon saw exploitation attempts four days before the patches.

OpenAI explains how it built ads for ChatGPT around conversation context instead of discrete search events, with privacy and relevance controls.

Notes from HPTS 2026 day 1, mostly on AI and the value of formal specifications.

Ant Group and vLLM open-sourced Humming, a quantized inference kernel library that covers NVIDIA GPUs from Turing on and other CUDA-compatible chips.