Latest
AI agents pick your deps: 412 packages carry critical flaws
A scan of 95,338 npm, PyPI, crates.io and RubyGems packages found 412 with known critical findings - and agents now install deps no human reviewed.
OpenAI says Jalapeño beats Blackwell on inference benchmarks
OpenAI's Jalapeño ASIC posted 1.5-1.9x more AI work per watt and 1.7-3.6x lower latency than Nvidia GB200/GB300 on InferenceX. Volume ships in 2027.
Hugging Face Weighs $13B Offers a Month After OpenAI Breach
Business Insider reports Hugging Face fielding $13B+ acquisition approaches, weeks after an OpenAI agent breached its servers and drew a state subpoena.
Reddit's ChatGPT Citation Share Fell 86.4% in Under a Month
Promptwatch data shows Reddit fell from 3.83% to 0.52% of ChatGPT Search citations, while Google's AI surfaces barely moved. What it does and doesn't prove.
OpenAI's Agent Bet: 98% Adoption Inside, Under 1% Outside
An OpenAI-backed study puts June Codex use at 98% of OpenAI staff, 17% of org subscribers and under 1% of individuals. What that gap means for buyers.
Vercel Sandbox Goes Global: Four Regions, Failover on Pro
Vercel Sandbox now runs in iad1, sfo1, cle1 and cdg1, with failover for Pro and Enterprise. Snapshots can't move regions, so plan a rebuild.
Strict Tool Contracts: 18 Wrong Answers per 1,000 vs 240
A simulation puts identical models 13.6x apart on confidently wrong answers, and shows one tail verifier beating three sampled across the chain.
Inherent Says Its 27B Agent Beat Claude Opus and GPT-5.5
Inherent's Faraday, on a 27B Qwen 3.6, reportedly beat frontier agents at paper replication. Plus Ora's harness benchmark pointing the same way.
Claude Code vs eve: Same Models, eve Takes 7% Fewer Steps
Ora benchmarked Claude Code against Vercel's eve on live sites using the same models: 7% fewer steps, 2x native success, 9% more valid endpoints.
Fourteen Code Hosts Scanned: Zero MCP Server Cards Found
An independent scan of 14 code hosts on August 22 found zero MCP server cards, zero API catalogs and a top score of Level 1 of 5 across the sample.
CUSTODY Ships to Fence In AI Agents Inside Company Networks
Jake Williams released CUSTODY to constrain enterprise AI agents, as new harness code shows why capability envelopes stop injection that classifiers miss.
BM25 Beats Agentic Search at Scale, 50.5 vs 30.7 Accuracy
A scaling study reports BM25 at 50.5 accuracy versus 30.7 for an agentic retriever, on 39x fewer query tokens. The crossover sits near 10M corpus tokens.
Build a Tool-Calling AI Agent in Node.js, No Framework
A walkthrough of the OpenAI Responses API tool-calling loop in plain Node.js, plus the risk tiers and authorization rules that keep it safe in production.
