---
title: "AI engineering, November 9, 2026"
description: "The Forward Pass daily issue, November 9, 2026."
canonical: https://theforwardpass.net/archive/daily/2026-11-09
updated: 2026-11-09
---

# AI engineering, November 9, 2026

The Forward Pass daily issue, November 9, 2026. Source: https://theforwardpass.net/archive/daily/2026-11-09

> This issue is researched and written by AI models, and every fact is checked against its cited source. No human edits it before it is sent.

_Apiiro finds Claude Code secrets, Anthropic secures gas-fired power, SweepANN cuts molecule-search time, Microsoft VP releases gputaskman_

SweepANN searched roughly 500 million molecule vectors in 61.04 seconds. It returned the top 500,000 candidates for each of more than 26,000 protein pockets.

Meanwhile, a Claude Code security report identified a mismatch between machine-local ignore rules and project files that travel with shared code. Anthropic struck a $3.4 billion deal to generate gas-fired power for data centres.

If you read one thing today: the Claude Code approval-file issue. It explains why the location of an ignore rule matters when your team shares project files.

## Apiiro finds secrets in 1,122 public Claude Code project files

**Top News** · 598 HN points

Claude Code's “don't ask again” option saves approved shell commands in a project file. Apiiro found that credentials embedded in those commands ended up in public copies of that file on GitHub.

Here's what happened:
- Approved shell commands were recorded in the project file, including commands containing credentials.
- Apiiro found 104,185 public copies of the file on GitHub.
- Of those copies, 1,122 contained a real secret, making this an observed credential exposure rather than just a possible failure mode.

One catch: the ignore rule is local to the developer's machine, while the file lives inside the project. That mismatch means the protection can fail when code is shared, because the project file and the ignore rule do not travel together.

Sources: [runtimewire.com](https://runtimewire.com/article/claude-code-dont-ask-again-secrets-github-apiiro)

## Anthropic strikes $3.4 billion deal for gas-fired data centre power

**Top News** · Reported · ft.com · 40 HN points

Anthropic is securing gas-fired power for its data centres through a $3.4 billion deal. The agreement is designed to bypass traditional suppliers and secure scarce equipment as energy demand rises.

Here's what the deal covers:
- Gas-fired generation is the source of the power Anthropic is arranging for its data centres.
- The $3.4 billion agreement gives Anthropic a way around traditional suppliers, rather than relying on that supply route.
- Access to scarce equipment is part of the goal, tying the power deal to the hardware needed to generate electricity.

Why care? Anthropic is bypassing traditional suppliers to secure scarce equipment amid rising energy demand. It has struck a $3.4 billion deal to generate gas-fired power for its data centres.

Sources: [ft.com](https://www.ft.com/content/41bf5c6a-9735-4f1a-a8e3-2bcccad58ba0?syn-25a6b1a6=1)

## SweepANN cuts day-long drug-screening retrieval to 61 seconds in tests

**Top News**

SweepANN accelerates the retrieval step used to find molecules for protein pockets, the binding sites a drug targets. Researchers report cutting a full query round from roughly 24 hours to about a minute.

Here's how it works:
- The reported run searched roughly 500 million molecule vectors for more than 26,000 protein pockets in 61.04 seconds.
- Each pocket received the top 500,000 candidates, with 99.63% average recall across the run.
- Batches of 64 pockets share each vector-shard read, with matrix multiplication for scoring and temporary thresholds to avoid sorting the full score matrix.
- INT8 quantization cuts storage from about 1.54 TB to roughly 387 GB across four RTX PRO 6000 Blackwell GPUs.

One catch: the 61-second run excludes about 502 seconds for quantization and loading. That setup cost can be spread across repeated searches when the molecular library does not need re-encoding.

Sources: [hyper.ai](https://hyper.ai/en/stories/73e168e3f3cf0e398eee435ff30abcbf)

## Signals

1. [Microsoft VP releases gputaskman, an open-source GPU monitor tracking per-process VRAM on Windows, Linux and macOS](https://hyper.ai/en/stories/4aec373456ad9ac9b9966fe843b17071) · Reported · hyper.ai
2. [MaxxBench tests whether coding agents follow changed requirements across 89 modified DeepSWE tasks](https://hyper.ai/en/stories/1dd53774aab2158446bbab6e74a0e1ba)
3. [Homelabber runs 2.8-trillion-parameter Kimi K3 CPU-only with Intel Optane PMem 200 at 0.85 tokens/second](https://www.tomshardware.com/tech-industry/artificial-intelligence/homelabber-uses-second-gen-optane-memory-sticks-to-run-a-2-8-trillion-parameter-ai-model-on-cpu-alone-warns-that-optanes-extreme-heat-can-burn-out-nearby-memory-dram-used-blower-fans-to-prevent-melting-ram)
4. [Tencent adds three video models to WorkRally, including an upscaler it says is 80% faster](https://pandaily.com/tencent-workrally-upgrade-rally-video-edit-upscale-models)
5. [SLA robot model beats π0.5 by 24.0 percentage points in average real-world success rate](https://hyper.ai/en/papers/2611.08373)
6. [Vercel’s agent-browser CLI gives AI agents browser control through accessibility trees and shell commands](https://github.com/vercel-labs/agent-browser) · 111 Stars today