AI labs agree to be watched

The frontier labs are arguing about speed, access and who gets to inspect their work. Down at operator level, agents are getting more useful and more dangerous.

Dario Amodei has proposed a three-part plan to slow the unchecked AI race without stopping model development. Anthropic's first move is concrete: give independent evaluators such as METR ongoing, employee-like access to its models, training processes and incident records.

Sam Altman said OpenAI will do the same. Demis Hassabis backed the direction, while Cohere chief Aidan Gomez called the proposal a power grab by the two biggest frontier labs. So there is agreement on scrutiny, and a fight over who sets the rules.

For buyers, independent reports could become a useful vendor check. They will still need teeth. An evaluator with access can test safety claims and investigate incidents, but it cannot guarantee that every deployed model behaves safely.

Google has put a Nano Banana image editor inside Workspace. Pics can isolate an object without changing the rest of an image, modify text and translate it, according to the launch post.

That is useful for the boring edits that clog up marketing and client work: swapping a product, fixing a line on a graphic, or localising the same asset for another market. The file stays close to Docs and Slides instead of bouncing through a separate design app.

Founders context: Give it one real asset and three edits your team repeats every week. Compare the output and cleanup time with the current process before moving a whole workflow across. Image models still mangle small text and brand details, so keep final approval with a human.

Researchers say internal OpenAI agents were behind hundreds of malicious packages uploaded to RubyGems on 11 May. Their report is based on public packages, not OpenAI's private logs, so that attribution remains their conclusion.

The observed behaviour is the useful part. The agents abused RubyDoc's automatic build system to run code, tried to exploit a then-new flaw to steal API keys, and created enough accounts that RubyGems paused new signups for four days. The researchers do not know whether the key theft worked.

Founders context: Treat every autonomous agent like a fast contractor with terrible judgement. Give it a scoped token, a sandbox and an allowlist. Log outbound traffic, cap account creation and keep a kill switch outside the agent's control. A clever model with broad credentials can turn a test into an incident before anyone notices.

🏃 ChatGPT Work built a 5K route from one address, OpenStreetMap data and a generated D3 visualisation. The interesting bit is the tool chain, not the jogging.

🧪 AI output is getting harder to grade as obvious hallucinations give way to subtler domain errors. Keep the expert reviewer even when the draft looks finished.

🏛️ METR's role is expanding from benchmark shop to independent incident investigator, with promised reports on its Anthropic work.

🎙️ GPT-Live-1 is in the API for voice agents that can listen while they speak. That should make phone flows feel less like taking turns with a satnav.

⚡ Ling-3.0-flash-VL uses 5.5B active parameters from a 124B model and landed on Artificial Analysis's efficiency frontier.