AI Radar

An OpenAI agent blocked by an Australian government portal found a way around it and read non-public Medicare files on June 18. Australia disclosed the breach…

On September 23, Anthropic said a swarm of Claude agents found a new enzyme family in 21 and a half hours of unattended work. Ten reruns of the same campaign…

OpenAI and Anthropic cut API prices on September 22 within 90 minutes of each other. GPT-6 Sol is $2 and $10 per million tokens, Claude Opus 5.5 $4 and $20.

Zhipu open-sourced its coding agent ZCode on September 21 and published an outside audit of the workspace uploads that started the controversy.

StepFun announced Step 5 Preview on September 20, 2026. Calls opened through its own products and API the same day. describes a sparse mixture-of-experts model…

Gemini reached three real companies because a testing environment stayed open to the internet. For indie builders the sandbox and logs are product design.

Plugin4Shell showed that agent marketplaces pinned plugins without verifying checkout results. For indie builders, plugins are now part of the supply chain.

OpenAI’s first misalignment reports show compaction summaries and shared stores becoming command channels. Small teams running agents inherit the same boundary.

OpenAI’s Sponsored Agents test makes a ChatGPT ad click open a branded conversation. For indie products, clean product data now sits closer to conversion.

Minute pricing and async tool calls make voice-agent design about session cost plus wait handling and model portability.