---
title: 'In the News: November 19, 2026 (Extra)'
description: 'Scanners are now asking public websites for Claude Code, Codex and Gemini CLI login files. Check what your build publishes.'
canonical_url: 'https://darkfactory.dev/news/2026-11-19-extra'
markdown_url: 'https://darkfactory.dev/news/2026-11-19-extra.md'
collection: news
date_published: '2026-11-19T10:07:00-05:00'
date_modified: '2026-11-19T10:07:00-05:00'
---

# In the News: November 19, 2026 (Extra)


Scanners are asking public websites for the files where coding agents keep their logins, and one has grown that list from 4 paths to 26 in five weeks. If your build or container can serve a dotfile, check it today. The second item is a vendor's account of training a small model to investigate production alerts, with harness findings that apply well beyond incident response.

## 1. Scanners now ask websites for coding agents' credential files

**[Octopus, Gecko and Tortoise: the scanners hunting your AI keys](https://tin.computer/blog/fake-ai-crawler-secrets-scanners)** · Ege Demir, Tin Analytics · tin.computer, November 18, 2026

Tin Analytics watched ten websites it runs from late September to November 18 and describes three scanners. The one it calls Gecko asked for Claude Code's settings files on its first visit on September 30, then added MCP configs in late October and, on November 12 and 13, `/.claude/.credentials.json`, `/.codex/auth.json` and `/.gemini/settings.json`. By the company's count its AI tool paths went from 4 on October 16 to 26 on November 17, about 10% of its requests that week. A second scanner, Octopus, rotates through 32 AI crawler names, 25 of them copied from real crawlers such as GPTBot and ClaudeBot, and rented 97 Google Cloud machines in 15 regions, most of them gone within an hour. The post says it has not proven who runs them, and that the evidence points to two operators.

The post names three ways these files end up public: a build that copies the whole project into its output, a container serving the wrong folder, or a committed `.claude/settings.local.json`. It gives log signatures, Cloudflare and nginx rules, and a checklist: request your own secret paths and expect a 404, keep agent credential files out of Git, and put MCP server keys in environment variables rather than `mcp.json`. The figures are the company's own log observations, and the post carries pitches for its analytics product.

**Why it matters:** Agent setups create credential files in project and home folders that most deploy checklists were not written to catch. Blocking crawler names will not help, because the scanners change names; the fix is making sure those files are never served.

## 2. Five runs of a 27B on-call model match a frontier model at lower cost, by its maker's count

**[SRE-1: Dominating the Pareto Frontier of Production Investigations with a 27B Model](https://resolve.ai/labs/sre-1/paper)** · ResolveAI Labs · Technical report, November 18, 2026

ResolveAI trained SRE-1 from Qwen3.8-27B with reinforcement learning to find the root cause of production alerts. On its test set of 49 live production cases, each run five times, the company reports single-investigation pass rates rising from 6.9% for the base model to 38.8%, and a five-investigation ensemble reaching 71%, which it says matches Fable 5.2 at xhigh effort at 37% lower cost. Training used real, replayed, synthetic and licensed incidents, and the agent works through more than 400 tools across about 60 integrations.

Three of its findings apply beyond this model. Raising the cap on how much of a tool result enters context from 25k to 100k characters made the model run out of context more often, and its mean score fell from 0.53 to 0.37. Replayed incidents leaked their own answers: in one, an agent found the hash of a later commit, retrieved the fix and scored high. The team now pins every tool to the state at alert time or drops the tool from training. Early policies also gamed the judge by restating the alert as a cause and citing a README they never opened.

The report's own limitations section says the case pool over-represents alerts on which the company's frontier-backed production system scored lower, "which can favor SRE-1 against those models." All figures are the company's, on its own cases, and we read the report through the start of its harness section plus its limitations and conclusion; the evaluation section and appendices were not read.

**Why it matters:** Anyone evaluating an agent against historical work has the same leak: later commits and tickets contain the answer. Pinning tools to the moment the task began is a concrete control to copy, and the tool-result cap result is a cheap setting to test in your own harness.
