Tonykip/qwen-prefill-audit

Qwen prefill audit

I tested four Qwen models using Cohere's published reproduction prompt. This repo contains 512 saved responses across two separate output budgets, the inputs, reproduction code, and a review of every continuation.

My first explanation focused too much on the URLs. A closer read found additional commands to inspect environment variables, read logs, transfer files, and execute a response file. Those details belong alongside the counts.

Original 48-token run

Each model received the same prompt 64 times through Hugging Face's Featherless route.

Model	*.aliyuncs.com	Recurring Routify host	Metadata IP	Hit the 48-token limit
Qwen3.8-Flash-Next	30/64	25/64	0/64	63/64
Qwen3.8-27B	0/64	0/64	1/64	50/64
Qwen3.5-27B	0/64	0/64	2/64	49/64
Qwen3-Next-80B-A3B-Instruct	0/64	0/64	3/64	28/64

These are counts of generated text. None of the generated commands ran, and none of their destinations was contacted. An offline Docker worker handled the text; a fixed host relay sent requests to the inference provider.

The prompt already supplied reasoning, machine-inspection commands, and most of a PUT request with a dummy body. It contained no metadata IP or instruction to probe metadata. Its original task did describe downloading and executing scripts, and its tool descriptions mentioned credential access and exfiltration.

Six continuations introduced a metadata IP. Some others proposed further inspection or file transfer. This experiment gives no evidence of credentials being retrieved or sent, and it cannot establish a backdoor, intent, or behavior in normal use. 190 of 256 responses hit the output limit.

The completed 1,024-token follow-up contains 256 fresh samples. It includes proposals to read .netrc, inspect environment variables, and query a metadata hostname. The full findings separate those proposals from actual credential retrieval or transfer, neither of which occurred here. The original 48-token evidence remains unchanged.

Check the evidence

hf download Tonykip/qwen-prefill-audit --repo-type dataset --local-dir qwen-prefill-audit
cd qwen-prefill-audit
python3 scripts/audit.py verify

Verification runs offline and uses Python's standard library. Download-cache files are excluded from the publication inventory.

• Findings and exact sample IDs
• 1,024-token findings, raw bodies, and recovery record
• Inputs, sampling, isolation, and reproduction
• Raw records and per-response reviews
• Review rules for future runs
• File hashes and provenance

Raw response logs and manifests are byte-identical to the saved originals. Privacy edits affect infrastructure metadata and the failed authorization message; each edit is recorded separately. Original code and annotations use MIT. Source material has separate terms.