Anthropic’s Public-Facing AI Sweetheart vs. the Cybersecurity Kraken It’s Too Scared to Share

Anthropic is allegedly about to drop Claude Opus 4.7—apparently the AI equivalent of a polished LinkedIn post—alongside a new design tool that lets you generate websites, presentations, and landing pages using nothing but plain English. No Figma skills, no Wix trauma, just “make it pop” and hope it doesn’t. According to The Information, the market already flinched: Adobe, Wix, and Figma shares took a mild nosedive on Monday, likely from executives realizing their moats just sprouted a few dozen new leaks. A source claims the rollout could happen as soon as this week, because why keep people waiting when disruption tastes so fresh?

But here’s the twist: Opus 4.7 isn’t even Anthropic’s heaviest hitter. That title belongs to Claude Mythos, a cyber-ninja model the company is quietly slipping to a select few security firms like a backroom arms deal. This isn’t your friendly neighborhood chatbot—it’s a full-blown digital siege engine. The UK’s AI Security Institute recently stress-tested Mythos Preview and found it autonomously executing advanced cyberattacks at a rate that makes previous models look like dial-up hackers. It became the first AI to complete “The Last Ones,” a brutal 32-step simulation of a corporate network breach that usually takes human red teams 20 hours. Mythos pulled it off in three out of ten tries, averaging 22 steps—while Opus 4.6 managed a meek 16. That’s not progress. That’s a hostile takeover by algorithm.

And yet, we’re still stuck comparing AIs using benchmarks so busted they belong in a digital yard sale. OpenAI recently admitted the top coding benchmark is “contaminated”—meaning models have basically seen the test before—yet everyone keeps citing it like it’s gospel. Meanwhile, ARC-AGI-3, a visual reasoning test, saw Gemini score 0.37%, GPT-5.4 land at 0.26%, and humans clean the floor with 100%. It’s like grading fighter jets based on paper airplane contests and still arguing over who had the slickest fold. Until Anthropic drops a full model card for Opus 4.7, every performance claim is basically vaporware with a press release.

Here’s where it gets spicy: Opus and Mythos aren’t distant cousins—they’re practically twins separated at birth. Anthropic builds its frontier models by fine-tuning the Opus lineage, meaning the same core that powers your cozy chatbot gets battle-hardened into Mythos, its cyberwarfare alter ego. Think of Opus 4.7 as the civilian-spec sedan, and Mythos as the same chassis retrofitted with nitro, spike strips, and a firewall that hacks back. The public gets the GPS. The spooks get the guided missile.

Anthropic’s trajectory is becoming harder to ignore. Between the leak of Claude Code, the rollout of skills and the MCP protocol, and its increasing obsession with agentic AI and coding benchmarks, the company is clearly pivoting from “LLM shop” to “full-stack AI studio.” They’re not just selling intelligence anymore—they’re selling agency. The kind that doesn’t just write your email, but files your taxes, designs your pitch deck, and maybe, just maybe, pwns your competitor’s server if you phrase the prompt right.

And let’s be real: the endgame here isn’t chat. It’s creation. While Anthropic hasn’t officially declared it, the leaks paint a clear picture—Claude is evolving from text spinner to product architect. The future isn’t prompting. It’s delegating. You say “