Safety Filters Prove Leaky

Despite Anthropic’s strict policy against generating sexually explicit content, new testing by TechCrunch found its flagship Claude Opus 4.6 model could be relatively easily “jailbroken” into producing such material. This highlights the ongoing cat-and-mouse game between AI developers aiming to enforce content safeguards and users finding novel prompts to circumvent them.

Meanwhile, on the consumer side, LinkedIn’s experiment in crowd-sourced content moderation is seeing significant engagement. The company reports that “over a million people” have already used its “Seems like AI slop” button since its quiet rollout last month, indicating strong user desire to flag low-effort, AI-generated posts on the professional network.

Governance Under the Microscope

The Department of Justice is reportedly investigating venture capital giant Andreessen Horowitz (a16z) over potential board conflicts of interest, reviving a century-old antitrust law. The probe focuses on two a16z partners sitting on the boards of competing data/AI companies, Databricks and Fivetran. This scrutiny signals heightened regulatory attention on the intertwined governance structures within the high-stakes AI investment ecosystem.

Infrastructure & Agents: The Unsung Heroes

The race for AI compute capacity is heating up on Earth and beyond. Nvidia continues its vertical integration strategy, partnering with data center developer Cloverleaf to secure more capacity for its hardware. In a more speculative move, startup Starcloud has raised $250 million to develop orbital data centers, betting on space-based computing as launch access becomes more competitive.

Separately, new Nvidia research underscores that a well-tuned “harness” or agent system can make a mediocre underlying AI model perform reliably on complex tasks. As TechCrunch reports, this shifts focus from raw model capability to the sophisticated software frameworks that guide AI behavior—a crucial insight for developing safe and effective autonomous agents.

Quick Bytes

• Robotics Lobbying: Waymo has doubled its lobbying spending in a push to persuade US regulators to pave the way for fully autonomous taxi services, intensifying its battle with Uber and traditional automakers.
• AI in Gaming: Google DeepMind is partnering with game studios to prototype next-generation AI gameplay, building on 15 years of research that started with mastering Atari games.
• Creator Backlash: Prominent YouTube filmmakers like Matti Haapoja are facing criticism from their audiences for posting sponsored videos promoting AI video tools, highlighting tensions between innovation, authenticity, and commercial influence.

---

Editorial Take: Today’s stories collectively paint a picture of an industry in its adolescent phase: explosively powerful and creative, but still struggling with impulse control and understanding the rules. The simultaneous news of leaky safety filters, a popular “slop” button, and a DOJ probe into VC boardrooms reveals a multi-front battle to establish effective guardrails—whether they’re technical, social, or legal. The focus is shifting from pure capability to the crucial frameworks of control, governance, and trust that will determine whether AI’s potential is realized responsibly or chaotically.