Anthropic Just Gave Defenders a Hacker's Best Tool

Morning, folks. Today's a mixed bag: Anthropic's handing defenders a hacker's own tool, Google's letting a rival sit at its table, DeepMind's trading games for a 20-year-old virtual economy, Alibaba's $10B bet got punished by its own investors, and Anthropic just admitted writing code was never the hard part. Let's get into it.

Anthropic is putting Mythos 5, its most capable cybersecurity model, to work inside Claude Security, live now for every Enterprise customer. Nobody actually touches the model, though; you hand it a repo, and it hands back a vulnerability, a severity score, and a suggested patch, no prompt box, no direct access. Mythos 5 itself stays locked behind Project Glasswing either way.

The model's already found over 10,000 high and critical vulnerabilities across major software worldwide through Anthropic's existing partners. Alongside the launch, Anthropic's also putting $35 million into open-source security work through a new fund. The logic here is simple, give defenders the results, keep the actual weapon locked away.

SpaceXAI just brought Grok 4.6 into Google's Enterprise Agent Platform, so any company already on Google Cloud can use it without a separate xAI account or billing setup. It's built for long coding and research jobs, with a 500,000-token context window and four reasoning levels from low to high, priced at 2/6 per million tokens.

What's actually notable is Google's own move here; Grok now sits in Model Garden right next to Gemini, alongside Anthropic's Claude too. Google's choosing to be the neutral marketplace even for its most direct competitors, rather than keeping customers locked to its own model.

Google DeepMind took a minority stake in Fenris Creations, the studio behind EVE Online, as part of a $120 million buyout that split the studio from its old publisher. The goal isn't building a better game AI, it's stress-testing whether AI can remember, adapt, and plan over timescales no other environment offers, inside a persistent economy that's been running nonstop since 2003.

DeepMind's own SIMA 2 agent already handles plenty of games it was never trained on but still struggles hard with anything long-horizon, exactly EVE's specialty. For now, testing happens in a separate offline version, not alongside real players. Fenris's CEO put it simply: AI needs to learn on timescales "no other game environment demands."

Alibaba is selling 710 million new shares in Hong Kong to raise $10.2 billion, all of it going straight into AI, chips, infrastructure, and model development. It's the largest follow-on share sale in Hong Kong's history. This comes on top of $10 billion in capex last quarter alone, up 75% year over year, and a $56 billion three-year AI commitment already in motion.

The market's reaction says a lot, though; Alibaba's stock dropped as much as 10% on the news, its steepest fall since April 2025. Profit is down more than 75% this quarter, and the company burned through $6.6 billion in free cash. Investors clearly aren't convinced the spending's paying off yet, they're just watching it get bigger.

Anthropic published a new playbook showing Claude now writes roughly 80% of the code merged into its own codebase; engineers ship 8x more per quarter than they used to. But writing the code was never really the slow part; instead of handing engineers a long spec, Claude turns a request into a plain-language intent.md file a product owner actually reads first before anything gets built.

That's the real shift. Planning, testing, approvals, these can still eat days or weeks even when the code itself takes minutes. Anthropic says some planning work that used to take weeks now takes hours instead. Speed didn't move to the coding. It moved to everything sitting around it.