Latest
AI Code Passes Tests but Only 18.4% of It Is Secure
A 145-CVE benchmark found Claude Code passed 74.0% of functional tests but wrote secure code on just 18.4%. Naming the safeguard moved the number.
Parallax beta.8 Shows a Real iPhone — Agents Can't Tap It
Parallax now streams a plugged-in iPhone beside a Galaxy S24, and beta.7 cut agent taps from 14s to under a second on Android hardware. Here is the catch.
Kimi K3.1 Pro Weights Are Open. Can You Actually Run It?
Moonshot opened Kimi K3.1 Pro's weights: 2.8T parameters, a 1.56 TB download, and 8+ B300-class GPUs to self-host. The hosted API is still the real route.
GPT-6 Cyber Hits Critical as Agent Shell Gates Still Leak
OpenAI's GPT-6 Cyber scores 48.5% on ExploitGym and is rated Critical. Meanwhile a newline slipped past an agent framework's bash gate 40 times in 70.
Gemini for Home Lands in AT&T's $30-a-Month Security Plan
AT&T's Connected Life now runs Gemini for Home: camera event descriptions, natural-language video search and daily summaries, gated behind $29.99/month.
AI Agents Booked Just 4 Uber Rides All Last Quarter
Uber disclosed that third-party AI agents arranged four rides in Q3 2026 and around 500 Eats orders in one week, out of 320 million total trips.
Anthropic's FlashProve Aims to Prove Which Model Answered
Anthropic published FlashProve, an open-source system for cryptographic proof that an output came from a specific model. No lab plans to deploy it yet.
OpenAI Says ChatGPT Search Is Now 20% of Google's Volume
OpenAI puts ChatGPT search volume at 20 percent of Google's. The definition is OpenAI's own and the Google figure went unsourced at the event.
Anthropic's New S-1 Filing: Compute Took 58% of Revenue
Anthropic's updated S-1 shows compute at 58% of revenue, up from 48%, with revenue per compute dollar falling from just over 2x to 1.72x on a $10B bill.
Agents Overreach on One Task in Five, a Benchmark Finds
AuthorizationBench scored 6,864 agent trajectories: 19.2% took an unauthorized action, and completing the task predicted nothing about staying in bounds.
Claude Message Threads Beta Cuts Request Size, Not Your Bill
Message Threads is in public beta on the Claude Messages API. Anthropic reports 35% lower time to first token — token charges stay unchanged.
Laya-MLX Claims 7-14ms Typed Decisions on Apple Silicon
A new MLX runtime says typed decision models route agent queries in 7-14 ms on an M3 Max. Every number is the author's, with no independent test yet.
Agents Score 12% on GDPx: Which Model Should You Pick?
SemiAnalysis's GDPx gives the best agent 12.0% on 206 real work tasks. The leaderboard is spiky, Carbon is far cheaper per point, and 98% broke rules.
