Ex-Anthropic lab Mirendil is in talks for $1B at a $5B valuation
Two stories worth your time: a self-improving-AI lab that quintupled its own price in three months without shipping a model, and malware that now settles its next move by a vote of four commercial LLMs.
Mirendil, the lab founded by former Anthropic researchers to build self-improving AI, is in talks to raise about $1 billion at a $5 billion valuation, with Kleiner Perkins negotiating to lead the round, according to Bloomberg. That is five times the price Mirendil set in late June, when it raised a $200 million seed led by Andreessen Horowitz and Kleiner Perkins with Nvidia participating. The company, founded earlier this year by Behnam Neyshabur and Harsh Mehta, now has more than 20 employees and expects to launch its first frontier model by early next year — meaning the new valuation is being assigned before the product exists.
The thesis is recursive self-improvement: models trained specifically to do AI research well enough to improve the next generation with little or no human help. Read that against the rest of this week. OpenAI published global-standards language saying fully autonomous self-improvement "should not be pursued unless and until it can be done safely" — we covered that in OpenAI wants Washington to lead on global AI standards — and the pressure to go further is already public: Ezra Klein calls for a ban on self-improving AI. The private market is answering the question on its own schedule, and the answer is a higher number.
That is the real signal here. A pre-product lab moving from $1 billion to $5 billion in a quarter is not being valued on revenue; it is being valued on the credibility of a research direction that regulators and even the labs themselves describe as the thing to slow down. Mirendil's own capital is going into compute rather than headcount — it signed a multi-year Google Cloud deal worth more than $100 million in August, roughly half its seed round, spanning both TPUs and Nvidia hardware. Investors are not just funding a research bet; they are funding the ability to run experiments at a frequency that a smaller budget cannot match.
Cisco Talos has documented what it says is the first publicly known Windows implant that delegates tactical command-and-control decisions to a panel of large language models. The malware, named CLOSEDQUORUM, is a 16.4MB Go-compiled executable that queries up to four commercial providers — DeepSeek, Qwen, Mistral and Google Gemini — asking each to choose the next action from a fixed menu: steal credentials, inject code, or establish persistence. Each model returns a vote, the majority wins, and ties break deterministically in a fixed order that gives DeepSeek the deciding say. There is no attacker-run server to block and no operator to wait for; the LLM endpoints are the command infrastructure, and stolen material leaves via an encrypted Discord webhook.
Talos's static analysis confirmed the decision loop, though the copy it examined shipped with placeholder API keys and a dummy webhook, so the researchers never watched it run end to end. The significance is architectural rather than operational — the human is displaced from a phase of the attack, not just made faster, and that phase no longer pauses when the attacker sleeps. The same publication also released CAIRN, an open-source toolkit for classifying AI-integrated malware by its "cognitive artifacts" — embedded prompts, provider endpoints, orchestration logic — after finding that an AI-specific evasion trick had spread from a red-team demo into unrelated malware within 12 months.
Why it matters: the failure modes Talos lists — provider refusals, rate limits, malformed output, predictable tie-breaking — are also the detection surface. Legitimate software contacts one of those providers; very little contacts several while touching credential stores and setting up persistence. Behavioral signatures, not domain blocking, are what defenders can actually use.
What to watch: whether Mirendil's first model lands on its stated early-2027 timeline, and whether anyone builds a second implant on this pattern now that the blueprint is public.
Would you treat "a model that researches AI" as a capability worth regulating before it ships? Tell us in the comments.
