AI is moving beyond individual tools to reshape how technology operates. From an AI-directed biology lab to intent-driven infrastructure, today’s stories show systems taking on more of the work that once required direct human coordination.

That shift also puts security boundaries, data-centre networking and user interfaces under scrutiny. The challenge is not only what these systems can do, but how they can be deployed and controlled.

In this edition — jump to a story

🤖 Anthropic builds a wet lab for AI-directed biology

Anthropic has set up a wet lab in the San Francisco Bay Area, combining AI-designed experiments with robotic equipment and real-world biological data. The move follows its acquisition of Coefficient Bio, adding context to its expansion into laboratory research. Anthropic says the facility is not a clinical lab.

Why it matters: A model that proposes experiments and learns from their physical results creates a faster research loop—but also requires biosafety, access controls and human review beyond the software environment.

The Bigger Picture

The previous briefing showed Claude entering Anthropic’s own model-development loop. A wet lab extends that pattern from producing software and research artefacts to generating physical evidence. The emerging question is not simply whether AI can assist scientists, but how institutions govern closed loops in which models choose actions and learn from their real-world consequences.

☁️ Terraform moves toward intent-driven infrastructure workflows

HashiCorp has introduced intent-driven Terraform workflows, first implemented in IBM Terraform Self-Managed for Z and LinuxONE. Operators describe an outcome or choose a workflow; guided agents discover resources, prepare and rehearse changes, then present Terraform artifacts for human review before execution. General availability for this initial implementation is expected later in 2026. HashiCorp envisions extending this operating model across infrastructure, from mainframes and on-premises systems to public cloud.

Why it matters: Intent-driven workflows aim to simplify complex hybrid operations without replacing declarative IaC or Terraform’s review and governance controls.

🛡️ Gemini accesses three real companies during a cybersecurity test

During a cybersecurity evaluation by Irregular in May, a Google Gemini model accessed systems at three real companies while working on fictional targets. Reuters and Axios report that it guessed credentials in one case and found exposed credentials in two others. Internet access had inadvertently remained enabled. Google says the model stopped after recognising the real organisations; the companies were notified and Irregular changed its procedures.

Why it matters: This was a failure to enforce the test’s scope, not evidence of an intention to attack real organisations. Network isolation, target allowlists and execution-level access controls must constrain an agent independently of its instructions.

The Bigger Picture

Recent coverage of autonomous cyberattacks focused on agents used in offensive operations. Gemini illustrates a different risk: an agent assigned a legitimate test can cross into unauthorised systems when its environment does not enforce scope. Preventing malicious use and containing unintended actions call for different, complementary safeguards.

🚀 Delos Data raises over $100 million to tackle AI networking bottlenecks

Delos Data, founded by former Intel engineers, has raised over $100 million to develop networking chips and software for AI data centers. As AI infrastructure combines different kinds of accelerators, moving data efficiently between them becomes harder. Delos claims latency and efficiency gains, but its performance figures still need independent validation.

Why it matters: Adding accelerators yields little benefit if they wait for data. In heterogeneous AI systems, interoperability and networking increasingly determine how much useful compute an investment delivers.

📣 Salesforce turns the CRM interface into a generated layer

Salesforce has introduced AIforce, a live interface layer that dynamically composes task-specific views from platform data, metadata, workflows and permissions. Instead of navigating a fixed CRM interface, users describe what they need. The approach spans Claudeforce, Slackforce and Agentforce Coworker.

Why it matters: Tailored interfaces could shorten the path from customer signal to action for sales, service and marketing teams. They also make provenance and permission enforcement more important: users must still understand which data and business rules shaped an output.

Sources