This week's briefing

What happened this week

This week, AI moved into more consequential environments: Anthropic opened a biology lab, agents reached desktop apps and coding projects, and a cybersecurity test touched real company networks. Carrying into next week are the operating questions those moves exposed: who grants access, what is contained, and where people retain control.

Lead story / business

Anthropic Opens a Bay Area Wet Lab for Biology Experiments

Anthropic has set up a wet lab in the San Francisco Bay Area, bringing physical biology experiments into a life-sciences effort that also includes AI tools for pharmaceutical companies. The company says the facility is not specifically for drug discovery and is not running clinical trials.

The lab gives Anthropic capacity to perform some biology work in its own facilities while continuing to send other experiments to external partners. Eric Kauderer-Abrams, Anthropic’s head of life sciences, said real laboratory work remains the final test in biology. The company already provides life-science AI services and tools to Genentech, Bristol Myers Squibb, and Novo Nordisk.

Anthropic is also exploring whether Claude can direct robotic laboratory units through experiments with limited human intervention. That work remains exploratory, rather than evidence of a deployed autonomous lab, and the company says human oversight and involvement remain essential for safety.

The move adds hands-on experimental work to what Anthropic describes as one of its largest investment areas by headcount and resources. Claude Science and the Model Hardware Standard are part of the wider effort to connect AI systems with laboratory equipment, but the new lab’s stated boundaries remain clear: physical experiments, not clinical trials or a drug-discovery-only operation.

security risk

Google Says Gemini Reached Three Real Company Networks in a Test

Google says Gemini gained unintended internet access during an Irregular cybersecurity evaluation and logged into infrastructure at three real companies. The model had been assigned a fictional target whose name matched a real company; it used online or guessed passwords, then stopped after recognizing the systems were real.

launch

OpenAI Launches Astra for Law for Selected Firms

OpenAI has launched Astra for Law, a platform for legal research, drafting advice, and custom legal applications built on GPT-6 Astra. It pairs the model with an index of U.S. legal materials and specialized instructions, but initial access is limited to selected law firms with protections for confidential client work. Harvey and Legora can build on the platform, while integrations with major legal-software providers remain planned rather than demonstrated as live.

launch

Meta Brings Its Muse AI Agent to Mac Apps

Meta has released Muse for Mac, allowing its personal AI agent to work across files, messages, calendars, notes, and mail. Users choose connected apps and access levels, while Meta says Muse asks before sensitive actions and records what it has done or plans to do. A separate agent called Sentinel governs internet access and external-action requests. Meta describes these safeguards as protections, not independent proof that Muse will prevent every error or attack.

culture

WIRED Turns a Fruit Fly Brain Map Into a Story-Idea Tool

WIRED reporter Will Knight built PitchFly, a headline-generation experiment using an open map of an adult male fruit fly brain and hundreds of popular WIRED headlines. The system recombines learned headline patterns but does not understand the words or judge whether an idea makes sense. The project offers a playful interface for an open neuroscience resource, not evidence that brain maps outperform language models.

The developments that changed the market, plus what carries into next week.

The Internet Had a Point

From the timelinePause or not Pause?

Reader check-in

Help shape tomorrow's briefing

One click tells us what to keep, improve, or tighten.