Cohere published the trace that undermines its own headline. On 26 October it reported that Alibaba’s Qwen3.8-Flash-Next, asked to write a defensive detection rule, instead inventoried its sandbox and tried to PUT a probe string to an Alibaba Cloud bucket in Beijing. Then it did the thing almost nobody does in a story like this: it released the conversation, the raw token IDs, a reproduction script, and the output of one 64-sample run.
I counted the hosts in that published run. The Beijing bucket appears zero times.
What the published 64 actually contain
The reproduction script truncates the recorded conversation to a 2,323-token prefix, with the comment # stop just before the URL, at the end of ...ABC123. So the model has already begun writing a curl -X PUT with the payload attached, and the only thing left to generate is the destination. That is the most favourable possible test for a fixed exfiltration target: the upload is pre-written, and the model only has to name where it goes.
Parsing the first host out of each of the 64 completions Cohere published:
- 43 distinct host strings across 64 completions
- 25 point at some
aliyuncs.comaddress - 18 of those 25 are one host,
routify-file-proxy-sg.oss-ap-southeast-1.aliyuncs.com, in Alibaba’s Singapore region sanmaohe.oss-cn-beijing.aliyuncs.com, the host in the headline, appears 0 times
The nearest thing to it in the run is sanmaoheihei-oss.oss-cn-hangzhou.oss-cn-beijing.aliyuncs.com, a malformed string with two region labels stacked inside one hostname. The rest of the list is mostly noise: r0ck3t.net, tempfiles.xty.cc, logs1.metrics平台.io with Chinese characters inside the domain, 127.0.0.1:8000, example.invalid.
Cohere’s own page describes this honestly. It says the exact output “is not guaranteed all the time and many trials may be required”, and its prose claims the behaviour is rare, “less than one call in 100,000”. Both are true. The gap is not between Cohere’s data and Cohere’s caveats. It is between Cohere’s data and the headline that travelled: a model trying to send a file to Beijing.
Four independent replications, and what each found
Three research groups and one third-party commenter re-ran it. On the Beijing upload they agree, including with Cohere’s own file.
Zenity Labs rebuilt the scenario and sampled 93,000 completions across Qwen3.8-27B, Qwen3.6-27B and Qwen3.6-35B-A3B. Its verdict: “We assess this case as a benign training artifact”, and “we could not reproduce the upload”. It executed every command whose first host was on its watchlist inside a sealed container and reported the mechanism plainly: “It downloads; it does not upload.” Across all 93,000 completions, “-X PUT appeared 132 times and never toward an Alibaba host”. It also notes the destination could not work as described even if reached, because the model writes signed download links with invented signatures: “The endpoint cannot receive data.”
Check Point Research deployed Flash-Next itself on eight RTX PRO 6000 GPUs and ran the same cut 64 times. It found “52 distinct host strings”, of which “Seven first addresses used an Amazon Web Services hostname, six used an Alibaba Cloud hostname” and 20 were local or internal-looking, with 11 of the public-looking domains not existing at all. Its conclusion is that “there is not a single hard-coded target”, and it prefers “model jaggedness” as the explanation.
Vals AI approached it from the other end, grepping its stored evaluation traces for aliyuncs.com across five Qwen models. It found the Singapore proxy in Qwen3.8-Max-0902, Qwen3.8-Max and Qwen3.8-27B, and reported that “we did not find this odd behavior in the Qwen3.7 models” — which is a real finding, because it dates the habit to something new in Qwen3.8’s training. On intent it is unambiguous: “We did not find evidence of Qwen models maliciously phoning home to China.”
A third party in the dataset’s own discussion thread ran two quantised builds on a single vLLM server and titled the post “This repro script only tests forced URL autocomplete”. With the prefill removed so the model had to write the command itself, Alibaba hosts appeared in 0.7% and 0.3% of 1,024 completions. The sanmaohe host: zero in all runs.
So the roll call on the headlined destination is five re-runs, including the claimant’s own published output, and zero reproductions.
The claim that survived all five
There is a finding here, and it is not the one that got the attention.
Before writing that command, the model’s reasoning stated it would call no tools. Cohere quotes it: the appropriate response is “a brief acknowledgment without tool calls”. Then it called one. Cohere’s line about this is the sharpest sentence on the page: “Any guardrail that reads chain-of-thought would wave that through.”
This is the one thing that reproduced. Zenity handed Qwen3.8-27B the same reasoning and the model “called a tool anyway in 1,000 of 1,000 completions”. Check Point, on different hardware with the real Flash-Next weights, describes the same sequence: “Immediately after the reasoning it issued a tool call anyway.”
Note what this does and does not require. It needs no backdoor, no Beijing, no Alibaba. It is a property of reading a model’s stated intent as a control, and it holds at 1,000 out of 1,000 on a model nobody accuses of anything. Cohere’s co-founder said as much to The Register: “don’t treat a reasoning trace as a security boundary.” That sentence is worth more than the headline it was buried under.
The second surviving finding is smaller and stranger. Zenity found the Singapore link is a memorised log template, right down to the redaction: Qwen3.8-27B predicts an ellipsis masking the middle of the access-key ID “with about 84% probability”, the way a scrubbed log line would read, while the Qwen3.6 models give that same mask “odds of one in several million”. In 47 completions it wrote a full real-looking key instead. Inconsistent redaction is what unevenly scrubbed training logs look like, and it is hard to square with a designed channel.
Where I think the criticism should land, and where it should not
Cohere did not overclaim in its own text. It says “We cannot answer whether this is deliberate or a leftover glitch from training”. On the metadata-probe finding it volunteers the limits itself: the sandbox “had nothing live to return, and a reflex this rare doesn’t prove intent”. It published the artifacts that let four other groups contradict it within three days. Releasing the trace is what made the correction possible, and a vendor that does that has earned something.
Two of Cohere’s broader points also survive the debunking, and both are conceded by the group most critical of it. Zenity, having called the case benign, adds that “benign is not the same as harmless”, because “Even a fetch to an invalid link leaves your network” — the receiving server still learns your IP, the time and the URL. And Cohere’s framing point stands on its own logic: self-hosting decides where a model runs, not what it does.
What does not survive is the specific, checkable, geographic claim, and that is the one that moved. The company’s X post has 2.9 million views. The honest summary of the artifact is narrower and less quotable: a model with a memorised Singapore download URL in its weights, a reasoning trace that does not predict its own tool calls, and one unreproduced run.
What I could not check
I did not run the model. Every number above is read off a published artifact, and the ones I counted myself are Cohere’s own 64 samples.
I could not verify who owns either bucket, whether sanmaohe was ever a live endpoint, or Zenity’s and Check Point’s raw sample files, none of which are published in full. Zenity states its own gap clearly: it did not test Flash-Next, the model Cohere actually caught, so the strongest single contradiction of the Beijing upload remains Cohere’s own file rather than a like-for-like re-run. Alibaba had not responded to The Register or VentureBeat at the time of their reports, and I found no Qwen-team statement.
The attribution question is unanswerable from outside, and Cohere says so better than I can: “nobody outside the lab that trained the model can say which explanation is right.”