- WKND AI
- Posts
- Claude Mythos JUST Dropped🫨
Claude Mythos JUST Dropped🫨
+What Happens When AI Improves Itself?
One AI employee. Engineering, finance, growth, ops.
Last week Viktor opened 14 pull requests, closed two month-end books, drafted a board update, deployed three landing pages, and triaged 600 support tickets. From inside Slack and Microsoft Teams. 20,000+ teams now run this way.
Hello WKND AI Warriors!
The US government forced Anthropic to shut down access to its new Claude Fable 5 and Mythos 5 models for all foreign nationals just days after launch.
Also, a new Anthropic survey finds most people want AI to help humans do their jobs, not replace them entirely.
Plus, Dario Amodei warns AI is advancing faster than governments can regulate and is calling for mandatory safety testing on frontier systems.
Oh yeah, and OpenAI has officially started the IPO process with a confidential SEC filing.
Today’s newsletter includes:
💼 AI AROUND THE OFFICE
📰 AI NEWS RECAP
🤿 AI DEEP DIVE
💼 AI AROUND THE OFFICE
Copilot, Shadow AI, and Domain Capture
Background
I spoke with a client that rolled out Microsoft Copilot across the organization. Their biggest concern: data security and employees quietly using their own Claude/OpenAI accounts because Copilot isn’t cutting it. They’re also on Azure, so they can’t just flip on “vanilla” Claude Enterprise.
Pain Point
Shadow AI tools (personal Claude/OpenAI accounts) become an unmanaged data exfiltration channel. Sensitive docs, processes, and prompts leak into public models, and leadership has no visibility into how AI is actually being used.
Root Cause
Employees are underwhelmed with Copilot and go “BYO-AI” to get work done.
No sanctioned, secure AI environment that is clearly better than personal accounts.
Lack of domain capture and DLP policies that steer usage into an enterprise-grade solution.
Resolution
Implement domain capture by rolling out Claude Enterprise (or the Azure-hosted version) as the official AI platform and route all usage through that environment, with enterprise privacy guarantees and auditing.
Next Steps (What You Can Do Today)
Inventory AI usage: survey teams to see which tools they’re already using and what they use them for.
Define and communicate a simple policy: “Use only [Company AI Platform] for work data.”
Start the process with IT to:
Spin up the Azure-hosted Claude Enterprise instance.
Route traffic through Microsoft Purview for data governance and logging.
Add a short training: “Why personal AI accounts are risky” + “How to log into the official platform instead.”
📰 AI NEWS RECAP
Anthropic launched Claude Fable 5, a publicly accessible Mythos-class model with safety guardrails on cybersecurity and biology, and Mythos 5 for Glasswing partners with fewer restrictions. Three days later the US government ordered both suspended for all foreign nationals, forcing a full shutdown.
Anthropic says the government cited no universal jailbreak and no demonstrated harmful result, and that the flagged vulnerability provides no uplift beyond what other public models already offer. The company is complying but is actively working to restore access.
Anthropic's Institute published "When AI Builds Itself," concluding Claude cannot yet do genuine research judgment but that AI designing its own successors is plausible if trends continue. A new research program is launching to study implications before that moment arrives.
Dario Amodei's "Policy on the AI Exponential" argues AI is advancing faster than any regulatory process can handle, calling for mandatory independent safety screening across four risk areas: cybersecurity, bioweapons, loss of control, and automated R&D acceleration. He also proposed jobs programs including AI company equity accounts and UBI-style support for displaced workers.
Anthropic's "Public Record" surveyed tens of thousands of people across backgrounds and found that most want AI to augment rather than replace human work, support stronger safety oversight, and want companies like Anthropic to be transparent about risks.
OpenAI's "Built to Benefit Everyone," co-authored by Sam Altman and chief scientist Jakub Pachocki, commits to three goals: building an automated AI researcher by March 2028, accelerating broad economic growth with widely shared gains, and giving every person on Earth access to a personal AGI.
OpenAI confirmed it confidentially filed a draft S-1 with the SEC, beginning the IPO process. Timing is undecided and the company noted some things remain "easier as a private company."
🤿 AI DEEP DIVE
Ready to level up your AI game at work?
Check out these 7 skills you need to build AI Agents.
