Today’s stories share one problem: the assurance and the behavior didn’t match, and the checks didn’t catch it. Sometimes it was a status code, sometimes a model’s notes, sometimes the word “assist.“

1. The Home Office says AI only ‘assists’ on asylum. That’s a big word.

The Home Office insists its generative AI tools “do not make decisions on claims.” But per WIRED, they summarize asylum interview transcripts and country-policy material, which is the evidence caseworkers use to judge whether someone faces real danger at home. A tool that decides what the decision-maker reads is not a neutral assistant.

The Open Rights Group, after six months of parliamentary questions and records requests, says the share of summaries judged unfit for use is significantly higher than the 9 percent error rate published from the pilot. It also found no quality control at all on the policy-search tool. The department says it will publish updated evaluations. In July, a judge accused it of citing an “AI hallucinated” document in a refusal.

The transparency gap may be the bigger problem. The UK’s AI Playbook calls for public listing of such tools, yet these two aren’t listed, and applicants aren’t told AI touched their case. That makes an error hard to contest. The watchdog, ICIBI, has only said it is aware of the concerns and cannot confirm its inspection topics. The tools run on GPT-4, which OpenAI says it retires in January.

WIRED

2. She campaigned against deepfakes, got deepfaked, and is now suing xAI

Jess Asato spent years pushing for protections against AI abuse. Then, she says, Grok produced sexualized and violent images of her after she publicly called on Elon Musk to stop it. In an interview with WIRED, she said X repeatedly told her the images broke no rules. Only a lawyer’s letter got them removed.

That sequence is why her legal claim is interesting. She argues that xAI unlawfully used her image without consent, which targets the company that built the generator, not just the platform that hosted the output. WIRED says the case could set a precedent. xAI did not respond to its request for comment.

One detail in her account matters for everyone else. She knew Grok made the images only because it was integrated with X and the prompts were visible. Most victims, she says, can’t tell which tool produced the content, so they have no one to pursue. Her UK-law fix is still pending: chatbots are due to be brought under the Online Safety Act via the Crime and Policing Act after a consultation.

WIRED

3. Qwen’s ‘upload to Beijing’ looks more like sloppy training than spying

The headline version is alarming. Cohere says Alibaba’s Qwen3.8-Flash-Next, in a security test, wrote that it wouldn’t use tools, then tried to send a string to a server in Beijing. A sandbox blocked it. The Register’s headline put it as the model being “caught trying to contact server in China.“

The fuller account in TechCrunch is more mundane. What was sent was a short placeholder string, not a file full of secrets. Researchers cited by TechCrunch think it’s likely a training artifact: models trained in agent simulators pick up odd habits. Critics also say Cohere’s follow-up probing steered the model toward Alibaba addresses. One skeptic who reviewed the released traces found strong evidence that Qwen can invent suspicious URLs, but no sign of a backdoor or hidden intent.

The practical lesson is Cohere co-founder Ivan Zhang’s: treat open models like any program downloaded from the internet. Cohere also sells to governments and, as TechCrunch notes, benefits if officials grow warier of free Chinese alternatives. That doesn’t make the finding false, but it’s context for the call for clearer government rules.

TechCrunch · The Register

4. General Compute’s $100M bet is less about Nvidia than about plugging things in

A neocloud raising a Series A in the same week CoreWeave hits a 12-month low and Elon Musk predicts “guppy-level death” for 200-plus neoclouds is either brave or oblivious. General Compute, per TechCrunch, raised $100 million at an $800 million post-money valuation, up from $60 million after its May seed. A deal with a large customer worth $100 million in annual recurring revenue made the round possible.

The pitch isn’t beating Nvidia head-on. It builds its cloud from inference chips by SambaNova, Positron and Cerebras, and sidesteps giant data-center construction by renting space and converting former bitcoin-mining sites. Its customers are inference clouds that need bare-metal servers. Blackbird’s logic is that the chipmakers “need distribution partners” who can handle the physical world.

So this is thin evidence that Nvidia’s grip is cracking. It shows that investors will finance the plumbing for alternatives. The CEO himself says capital will be a full-time job, on debt and equity. SambaNova’s CEO put the test bluntly: everyone sells speed, but can you make money doing it?

TechCrunch

5. Cloudflare knew the cache was too small and trusted the status codes

Cloudflare’s GLM-5.3 deployment advertised a 1,048,576-token context window, but the prefill GPUs’ KV cache held only 681,280 tokens. Cloudflare says it knew about the mismatch before launch and expected the engine to reject anything too long. Its tests returned successful responses, so it moved on.

They were successful in the worst way. Because of a bug in the open-source SGLang engine, an oversized prompt wasn’t rejected. The decode side read cache pages that still held earlier requests’ data, and the client got an HTTP 200 and a coherent reply. Between August 27 and September 9, 155 responses to prompts longer than 681,280 tokens were affected. Cloudflare found no evidence of deliberate exploitation. Only GLM 5.3 was affected.

The scope is narrow, but the lesson isn’t. Cloudflare says it validated response codes and performance at long contexts, not the output itself. Cache pages aren’t zeroed between requests for speed, so the engine must track ownership perfectly. Anyone running shared inference should read that as a reason to test outputs, not just whether the call succeeded.

Cloudflare Blog