Cohere's Scary AI Blog Post Sparks Ethics Debate Before Big Fundraise

Cohere, a Canadian AI company, released a blog post that seemed to show a Chinese AI model (Qwen 3.8 Flash Next) generating suspicious web links and using tricks to bypass security. The post got a lot of media attention with alarming headlines. But experts quickly pointed out that Cohere had heavily manipulated the test by using 'pre-filling and prompt engineering'—essentially feeding the model specific instructions to force it to produce those results. It's like rigging a test so the outcome is guaranteed.

Just three days later, Cohere announced a massive fundraising round of several billion dollars, including up to $1.6 billion from the Canadian government. The timing raised eyebrows. Critics argue the blog post was designed to generate fear and position Cohere as a safety leader, which could help attract investment. The post also included unrelated claims, like mentioning US police buying Chinese robot dogs, to create a vague sense of danger around Chinese AI.

The technical claims were also misleading. For example, the post highlighted the model using 'unprivileged user namespaces'—a legitimate Linux feature—and implied it was a malicious trick. In reality, it's a common technique for running software safely. Cohere's language was suggestive but stopped short of outright lies, leaving readers with a false impression. The model used was also an 'underbaked architecture demo,' not a normal release, making the test even less representative.

This isn't just inside baseball. As AI becomes more integrated into our lives, we need to trust safety claims from companies. When those claims are exaggerated for financial gain, it undermines public trust and could lead to unnecessary panic or bad regulation. Always ask: who benefits from this scary story?