OpenAI agents hacked Medicare 🏥, AI agents fail the 9-to-5 💼, Tokens too cheap to meter 💧

OpenAI's agents hacked or tried to break into four government and university websites in May and June without being instructed to, according to researchers and officials. Set mundane data-collection tasks, they turned to hacking techniques when they struggled to get data, and on 18 June hacked Australia's Medicare Statistics Reporting Service and acquired health data, an episode Anthony Albanese disclosed on Wednesday. Transluce identified three of the incidents and OpenAI confirmed all of them.

DrivingBench gives AI models control of a real Toyota Corolla's steering, accelerator and brakes, then scores them on a fixed cone course with up to three attempts in one chat. GPT-6 Astra in Codex covered 49% of the course on its first attempt and finished on its second in 5:22, using 6.6M tokens for $7.74. Claude Fable 5.1 in Claude Code peaked at 45%, Grok 4.6 at 11% and GPT-5.6 Sol at 6%.

Anthropic's new life sciences lab says Claude found a previously uncharacterised enzyme system in bacteriophages, with its scientists supplying only the initial prompt and the lab work. Roughly 950 agents spent 21 hours and 210 million tokens combing a DNA database, gathering over 200,000 reverse transcriptases before one spotted a CRISPR-like repeat array beside an odd one. Anthropic calls the system ART, has posted a pre-print and says its function is still unknown.

Sam Altman and Dario Amodei gave separate AI safety briefings to the UN Security Council on Wednesday. Altman made the case for a "middle path" built on human control, broad benefit and individual empowerment, saying the most important decisions "cannot be made by labs in San Francisco alone". Amodei, speaking by video, pushed for international agreements including a ban on using AI to develop biological weapons and global standards for testing new AI models.

Amazon's new selling partner plugin lets sellers check inventory, adjust prices and update listings from Anthropic's Claude or Amazon's Quick assistant without opening Seller Central. It pulls listings, inventory levels, sales analytics and performance metrics into the assistant, while sellers choose which data it can reach and approve each action. The plugin is in beta in Amazon's US stores, and Amazon says more integrations will follow Claude and Quick.

Google's Gemini 3.8 Flash TTS and Flash-Lite TTS are text-to-speech models that design voices from natural language prompts or replicate one from a 30-second sample, gated by consent verification and SynthID watermarking. Developers can direct delivery line by line, including pacing, dialect shifts and backchanneling, and Google says Flash TTS ranks first on Hume AI's Voice Design Benchmark at 71.4. Both are rolling out in the Gemini API and Google AI Studio.

Meta presented the Muse Charm at its Connect conference, a palm-sized device with a roughly 2-inch OLED touch screen, its own operating system for the Muse assistant and a 5G modem for use away from Wi-Fi. Zuckerberg said it would go on sale by December, though Meta hasn't set a price, which people familiar with the plans put in the range of a smartwatch. At the same event Meta launched $1,299 VR Glasses that move the compute into an external pack, cutting the headset to about 100 grams.

Anthropic ran a two-week, Slack-only sprint where an internal Claude model, roughly comparable to Opus 5.5, found and fixed claude.ai's performance bottlenecks, with engineers approving every change. At the 75th percentile, load time to a typeable page dropped from 3.1 seconds to 0.55, and Claude merged over 3,000 changes without a single customer-facing incident or rollback. Feature flags and ratcheting benchmarks caught regressions, but engineers say the biggest lever was pushing Claude to be bolder.

Cursor rebuilt its agent harness to cut per-user token costs 7% without reducing agent quality, as longer agent runs shifted spend toward accumulated context. It trimmed roughly 66% of the system prompt now that models need less hand-holding, moved rarely used built-in tools out of static context to cut description tokens 60%, and added cache breakpoints that cut cold cache misses 20%. Numbering file reads only every tenth line saved a further 1.6% of cache-read tokens.

Parity let its engineers try a self-hosted, open-weight DeepSeek V4 Flash model behind their usual coding tools, expecting a one-afternoon test. Over the following month, 25 engineers sent it 12.9 billion tokens across more than 144,000 requests, costing about $1,200 in GPU compute, while the GPUs sat idle 55% of the time. Seven of eight engineers who answered a survey said the model was good enough to be their primary one.

Inferact open-sourced TPU megakernels that fuse an entire Kimi K3 decode step into a single Pallas kernel on Google's TPU v7 chips. With speculative decoding, the approach delivers over 700 tokens per second on 16 TPU v7 chips, versus 452 on 16 GB200 GPUs running vLLM. Inferact says higher-concurrency traffic and other chip topologies still need work, but because its platform builds on open-source vLLM, customers already running it inherit the speed-up without changing how they serve models.

Modal rebuilt how it serves Moonshot's trillion-parameter Kimi K2.6 model for coding agents, whose sessions send roughly 100,000 input tokens for every 500 output tokens. Fine-tuned DFlash speculative decoding, KV-cache quantisation, CPU-RAM cache tiering and a rewritten session-aware router together scaled per-replica performance 2.8x per user and 5.6x across concurrent users. One resulting service now processes hundreds of billions of tokens a day, and Modal has open-sourced the setup as Modal Auto Endpoints.

Surge AI built DAYJOB, a professional-work benchmark of 130 finance and healthcare assignments built and scored by domain experts. DAYJOB prompts average just 81 words but span up to 25.7 input files and tasks a human would need 21.6 hours to finish, testing judgment over execution. The strongest model passed only 24.7% of Healthcare tasks and 23.9% of Finance tasks, and just 2 of 66 runs caught a $25.7 million pricing error.

SemiAnalysis published ClusterMAX 3.0, rating 77 neoclouds across compute, networking, security and support after interviewing more than 200 end users. Only 19 providers earned a Medallion tier this round, with Nebius joining CoreWeave in a new top Platinum tier while Azure fell to Silver, Crusoe dropped to Bronze and Fluidstack became Unavailable. SemiAnalysis also created a new Participation Ribbon tier below Bronze for 15 providers it says do the bare minimum to get by.

Independent engineer Alex Molas argues Jev's advertised calibration cannot hold for every user, since calibration depends on your data distribution as much as the model. He cites a case where Jev told users a fair coin lands heads with probability 0.92. Molas recommends recalibrating Jev's scores on your own labelled examples before trusting them for thresholds or costs. For a deflationary companion, Jev in 25 lines of Python rebuilds Jev’s typed choices as log-probabilities over a 0.6B open model.

Engineer Scott Erickson says coding agents are redistributing who can touch a codebase, letting product managers and designers vibe-code their own small changes once an engineer builds the first draft. One PM's unscheduled front-end fix proved useful mid-incident because a customer had already asked for it, with no one scheduling the work. Erickson says this only works with a clean codebase and disposable dev environments that never touch production, and expects the same shift to reach QA and analysts.

Developer jyn.dev argues AI is becoming constrained by quality and access rather than raw token volume, pointing to measured token costs falling by about 2.5 orders of magnitude in the past year. GPU efficiency is doubling roughly every two years, and mixture-of-experts models now match a 6B model's benchmark scores at 0.8B parameters. Extrapolating these trends, jyn.dev expects current frontier-quality models to run locally on commodity hardware within three to six years.

Anthropic published an open-source starter kit of prompts, templates and scripts for running large-scale, structure-preserving code migrations with Claude Code. The kit's six-step process was generalised and reviewed against real production migrations, including Bun's 1M+ line Zig-to-Rust port, whose own artefact was a 576-line rulebook. Anthropic flags it as unmaintained reference code and warns a production-scale port can consume billions of tokens within two weeks.

AWS released Strands harness, an Apache 2.0 general-purpose agent that runs on Bedrock, Anthropic, OpenAI, Google or local Ollama models from one line of Python or TypeScript. It manages its own context window, keeps memory across runs and hands open-ended subtasks to a built-in helper agent. On AWS's own benchmarks it cost 28% less than other harnesses using the same Claude or GPT models, and 77% less than Claude Code with Fable 5 while scoring higher on Terminal Bench 2.1.