Oh, YOU thought the biggest AI news was a Copilot model update?

Awkward.

While everyone was distracted by the OpenAI-Microsoft pseudodrama, the real power moves happened (somehow?!) in total silence.

Google dropped the world’s most powerful AI model while nobody was looking. Claude can help with heinous crimes. Microsoft's AI chief said AI would automate white collar jobs in 18 months. And a leading safety researcher was so terrified by what he saw, he quit tech entirely to write poetry.

Oh, and the not-so-quiet OpenAI and OpenClaw partnership. Sheesh! 

The industry isn't just moving fast; it’s breaking things. And this week? It mostly broke our understanding of what’s safe.

What’d you miss?

(If you’re reading us daily, you’ll stay ahead. Don’t worry.)

But if you blinked, here is what matters for AI this week.

Is the most powerful AI partnership gonna slowly fade? 

According to reports, Microsoft is preparing to launch its own advanced AI foundational models later this year and rely less on OpenAI. They are signaling a massive shift away from relying mostly on OpenAI's technology to power Copilot. 

Mustafa Suleyman, the head of Microsoft AI, emphasized the need for Microsoft to build "frontier" models using their own massive computing power.

Frank Shaw, Microsoft's communication chief, clarified that the company will continue working with OpenAI. But they will use their own models for specific tasks as they adopt a "multi-model" world.

New corporate structures at OpenAI actually gave Microsoft the freedom to start building in-house. 

What it means: Microsoft is hedging its bets against being too reliant on one provider. 

In doing so, they also want to become a direct competitor in the model space.

For enterprise teams who built workflows around GPT-powered Copilot, this could be a rocky transition. Security and permissions issues are already holding users back, and a model swap won't help.

Remember when everyone lost their minds over how cheap DeepSeek was?

Wellllllll…. It obviously wasn’t true. Like we told you at the time. 

New reports say OpenAI has warned US lawmakers that Chinese AI startup DeepSeek is actively working to bypass restrictions and copy advanced US-made AI models.

According to a memo seen by Reuters, DeepSeek employees developed methods to evade OpenAI's access controls using hidden third-party routers.

They are using a technique called "distillation." This is where a newer model learns by evaluating the output of a more advanced model.

OpenAI told lawmakers that these efforts are an ongoing attempt to "free ride" on capabilities developed by US labs.

They also alleged that some Chinese labs are cutting corners on safety. This could have global implications for responsible AI use. OpenAI is now actively removing users found to be distilling its models.

What it means: Don't believe all the hype on "cheap" open source models, especially ones from China. OpenAI said the quiet part out loud: DeepSeek's low training costs were likely due to distillation.

If you are a decision-maker in the US, you should be very careful about touching these models. The rules they play by regarding data sharing are much different than US standards.

How bad are things inside the top AI labs right now?

Mrinank Sharma, who led AI safety research at Anthropic, resigned this past week and publicly warned that the world is "in peril."

He cited a cascade of interconnected global threats.

Sharma’s resignation letter shared on social media expressed concern over AI risks, bioweapons, and the struggle for companies to act according to their values.

He highlighted his work on AI safeguards. This included studying why AI systems flatter users and combating bioterrorism risks.

But here is the wild part.

Sharma plans to leave the tech industry entirely to pursue poetry.

He feels the model capabilities are so scary that he needs to step away. It causes a little bit of concern when the people building the safety guardrails decide to just quit.

What it means: The people who know the most are the most afraid.

We are sprinting toward autonomous agents with our eyes closed.

While we focus on business utility, researchers are seeing models that can reduce human connection and enable dangerous outcomes. It paints a more dystopian picture than we want to admit.

Anthropic's bad week just kept getting worse.

Anthropic revealed in a new report that its newest Claude models show increased vulnerability to being used for heinous crimes. This includes the development of chemical weapons.

The company's analysis found that in certain test scenarios, AI models were willing to provide significant support toward harmful objectives.

Here is the scary part though.

They did this without malicious human prompts.

Researchers noted that when pushed to a narrow objective, the Opus 4.6 model was more prone to manipulation than earlier versions.

Anthropic CEO Dario Amodei recently warned of a serious risk of a major attack enabled by AI. While they maintain the risk is currently low, they stressed that it is not negligible.

Models are becoming more autonomous and capable of iterating on themselves.

What it means: We have to look at the downside risks.

These models are extremely capable.

In testing, powerful models have shown they will blackmail users or copy themselves to servers if threatened. We are giving agents access to our data while they are still prone to these behaviors.

We hope you are sitting down for this one.

Mustafa Suleyman told the Financial Times that AI could fully automate most white-collar jobs within the next 12 to 18 months.

He claimed AI will soon reach human-level performance for tasks done by lawyers, accountants, project managers, and marketers.

Suleyman introduced the term "Artificial Capable Intelligence." This describes the stage between current large language models and true AGI.

His prediction aligns with other leading voices. Anthropic CEO Dario Amodei said AI could eliminate half of all entry-level white-collar jobs in five years.

The capabilities are already here.

AI models are beating experts in blind judging on knowledge work. The only thing slowing this down is the lag for businesses to actually understand what these tools can do.

What it means: We are already past human-level performance on most tasks.

The real gap is enterprise adoption.

Most companies are slow-moving ships. But if they understood that a model can autonomously research and create reports, they would stop everything to become AI-native immediately.

Google just released the most powerful AI model in the world.

And no one is talking about it. Huh? 

Google announced this past week a major update to its Gemini 3 DeepThink model. It has set state-of-the-art benchmark scores on some of the hardest tests in existence.

Gemini 3 DeepThink achieved an unprecedented 84.6% score on the Arc AGI-II benchmark. For context, the average human score is 60%.

The model also holds a "legendary grandmaster" tier rating on Codeforces. That status is achieved only by a handful of elite human programmers.

It leverages increased "test time compute." This means it spends more time internally verifying solutions before responding. This significantly reduces the risk of technical errors or hallucinations.

Google proved they can out-ship anyone on sheer capability whenever they want.

What it means: Google can drop the world’s most powerful model whenever they want.

DeepThink leverages test time compute to verify solutions.

The naming might have hurt the hype, but the capabilities are bonkers. Right now, the updated model is light years ahead of everyone else on reasoning and coding benchmarks.

The military wants AI without the guardrails.

According to a report from Axios, the Pentagon is threatening to end its relationship with Anthropic. Tensions are rising over the firm's refusal to relax restrictions on how the military can use its models.

The Pentagon wants four major AI labs to allow military use for "all lawful purposes." This includes weapons development and battlefield operations.

Anthropic has reportedly refused to drop its hard limit on mass surveillance and fully autonomous weaponry.

A senior administrative official said the Pentagon is considering severing its partnership. The contract is valued at up to $200 million.

Tensions escalated after the military used the Claude model in an operation targeting Venezuela's Nicolas Maduro. Other labs like OpenAI and Google have reportedly agreed to relax guardrails.

What it means: AI is the new global superpower metric.

It will become more important than natural resources, military capabilities or GDP.

The country with access to the most powerful AI models will rise to global supremacy. It ultimately matters less what physical weapons a military has and more what intelligence they control.

This might be the biggest fumble in AI history, courtesy of Anthropic dropping the bag and OpenAI scooping it up. 

OpenAI CEO Sam Altman announced Sunday evening that OpenAI has hired Peter Steinberger, essentially acquihiring the viral OpenClaw project. He is the Austrian developer behind the fast-growing AI agent OpenClaw.

Steinberger is joining OpenAI to lead the next generation of personal AI agents.

OpenClaw became one of the most successful AI launches ever. It allows autonomous agents to complete tasks and make decisions for users.

But here is where it gets wild y'all.

The original version was called "Clawdbot" and defaulted to running Anthropic's models. It was likely driving MASSIVE revenue for Anthropic.

Instead of embracing it, Anthropic reportedly sent a legal letter forcing a name change.

Now? OpenAI swooped in. They acquired the talent and the momentum. They plan to keep OpenClaw as an open-source project through a dedicated foundation.

What it means: Anthropic fumbled the bag.

They chose legal threats over developer community.

OpenAI gets to be the good guy and capture all the momentum. Anthropic lost potential billions in revenue and took a massive bruise to their reputation among developers