AI Weekly: Claude Gets Copied, a New Agent Enters the Race — February 23-March 1, 2026

Two stories dominated global AI headlines this week, and together they say a lot about how fast the ground is shifting under both the big labs and the agent tools built on top of them.

Anthropic calls out DeepSeek, Moonshot and MiniMax for copying Claude

On 23 February, Anthropic published research accusing three Chinese AI labs — DeepSeek, Moonshot AI and MiniMax — of running industrial-scale "distillation" campaigns against Claude. In plain terms: they used roughly 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude, then used those outputs to train their own models faster and cheaper than building from scratch.

The scale varies by lab. MiniMax was the biggest offender, at over 13 million exchanges aimed at agentic coding and tool orchestration. Moonshot ran about 3.4 million exchanges targeting agentic reasoning and computer-use skills. DeepSeek's footprint was smaller, around 150,000 exchanges, focused on reasoning tasks and "censorship-safe" training data.

Anthropic says the labs used "hydra cluster" networks — thousands of fake accounts spreading traffic to dodge detection, sometimes mixing in unrelated queries as camouflage. In response, Anthropic has built new detection classifiers, tightened account verification for research and education tiers, and is sharing technical indicators with other labs and cloud providers.

Why this matters for SMEs: none of this changes what Claude can do for your business today. But it's a reminder that model quality gaps close faster than most people expect, and cheaper competitors sometimes get there by copying rather than innovating. If you're picking a model or vendor for automation work, that's a reason to weigh reliability, support and safety track record, not just benchmark scores or price.

Hermes Agent launches with persistent memory built in

Two days later, on 25 February, Nous Research quietly dropped Hermes Agent on GitHub: an open-source, self-hosted personal AI agent built to run persistently, remember across sessions, and improve itself over time by writing reusable "skill" files after it solves a task. Early adoption moved fast: 22,000 GitHub stars and 242 contributors within two weeks of launch.

That's a notable jump from the state of the art we covered a fortnight ago in our "Smarter Models, Riskier Agents" roundup, where agent frameworks were mostly competing on task completion and required manual setup to persist memory between sessions. Hermes bakes persistence in from day one instead of leaving it to the user to wire up.

It's too early to call this a mature category. But it's the same pattern we flagged in that earlier post: agent frameworks are now competing on autonomy and memory, not just task completion. Expect more "self-improving agent" launches through the year as labs race to define what a persistent AI employee actually looks like.

For Australian SMEs, the practical takeaway isn't "adopt this specific tool." It's that persistent, self-improving agents are moving from research demo to daily-use software very quickly, and the security model matters as much as the feature list. A framework built around a self-improving skills loop is, by design, one that keeps expanding what it can do on its own — which makes the permission boundaries around it more important, not less. Any framework promising an agent that "grows with you" deserves real scrutiny before it touches your business data or customer records.

If you're weighing up whether an agent framework, a custom Claude Code setup, or a simpler automation is the right fit for your business, that's exactly the kind of call we help clients make. Have a look at our Claude Code consulting work, or read how we think about saving real hours with Claude Code workflows. If you'd rather talk it through first, get in touch.