---
title: 'In the News: November 21, 2026 (Morning)'
description: 'HD Moore seeded about two million prompts with trackable secrets and watched AI providers, resellers and attackers use them.'
canonical_url: 'https://darkfactory.dev/news/2026-11-21-morning'
markdown_url: 'https://darkfactory.dev/news/2026-11-21-morning.md'
collection: news
date_published: '2026-11-21T07:01:00-05:00'
date_modified: '2026-11-21T07:01:00-05:00'
---

# In the News: November 21, 2026 (Morning)


HD Moore's team at Dragos put fake credentials and secret URLs into about two million prompts and then watched who used them. Attackers who had broken into at least three inference providers did, and so did marketplace resellers and several model providers, some of them on requests marked Zero Data Retention. Teams whose agents can read credentials and whose inference goes through a router should check where their prompts travel. Hugging Face also published a recipe that trains a small adapter for one specific coding harness, with large gains on hard benchmarks and a caveat about what the model learned.

## 1. Two million bait prompts show inference data leaking through routers, resellers and breached providers

**[Three breaches, endless leaks: Your AI supply chain needs canaries](https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens)** · HD Moore, Chief Research Officer, Dragos · Dragos blog, November 19, 2026

From July through November, Moore's team sent roughly two million prompts to AI services. Each prompt carried its own AWS key, DNS name, URL or other secret that had never been used anywhere else, and the team ran the servers those secrets pointed to. The post reports three kinds of leak. Attackers were stealing requests from three hosting providers, none of which knew until Dragos told them. In the largest case "a couple hundred" bait AWS keys were stolen, and the bait keys were used for tens of thousands of AWS API calls. Those calls listed S3 buckets, searched Secrets Manager and tried to enable Bedrock models. Marketplace sellers who undercut on price logged the prompts they routed, and individuals then tried the stolen passwords by hand. One Chinese AI lab, which the post does not name, fetched the secret URLs from more than half of the prompts sent to its models, including roughly two-thirds of those marked "no data collection" or Zero Data Retention. Prompts sent to US hosts that resell its models reached the same scraper.

The team also traced about two thousand routes, using image fetches and deliberately broken requests whose nested error messages reveal each hop. Several hundred included a system they could not identify, and one marketplace seller "quietly adds about a thousand hidden tokens to every prompt." Dragos names none of the providers involved. It says one has issued a preliminary notice and another has still not disclosed its breach. Separately, a Hacker News user posted what they describe as an email from OpenRouter saying a security researcher's test URLs, sent to MiniMax models from July to October, were later accessed by a third party. That email does not name the researcher, and the poster says OpenRouter has published no page or blog post about it.

**Why it matters:** Moore's own summary is that "a privacy flag only binds the first hop" and that "AI agents will happily send whatever they can read, including credentials." An agent that can read your `.env` file can put its contents into a prompt, and that prompt may pass through hops you never chose. The post's advice is cheap to act on: pick upstream providers yourself, stay out of reseller marketplaces, and seed a fake credential into the context your agents already send, treating any use of it as an incident.

## 2. Hugging Face trains one LoRA per coding harness and cuts tool calls by more than half

**[Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost](https://huggingface.co/blog/harness-adapters-at-scale)** · Lewis Tunstall, Nathan Habib and colleagues, Hugging Face · Hugging Face blog, November 20, 2026

The team froze Qwen3.8-27B and trained a separate LoRA adapter for each of three agent harnesses, pi, Hermes Agent and OpenCode, with reinforcement learning run inside the unmodified harness. A proxy between the harness and the inference server recorded the exact tokens the model produced. The tasks came from SWE-smith, SWE-Gym and pull requests turned into tasks with a tool called Tasksmith. Each task was checked so that its reference solution earns full reward and wrong-answer controls do not, and those controls caught a Terminal-Bench 2.1 verifier that accepted known-wrong answers. On the 51 hardest SWE-bench Pro tasks, the pi adapter took pass@1 from 39.2% to 64.7% (15 tasks newly solved, 2 lost), Hermes went from 29.4% to 43.1%, and OpenCode from 35.3% to 51.0%. On tasks both models solved, the adapters made 56 to 72% fewer tool calls on SWE-bench Pro and 53 to 70% fewer on Terminal-Bench 2.1. Each result is one attempt per task, and the authors warn that scores cannot be compared across harnesses.

The authors flag their own open question. With thinking turned off, the adapted model writes short text before tool calls that "can sometimes look like reasoning traces," which they call reasoning leakage. On 300 business-automation tasks run in a different agent loop, the adapters stayed within noise of the base model, except perhaps the Hermes adapter, which fell 5.3 points. The adapters, the task dataset and all evaluation trajectories are public.

**Why it matters:** If you run an open-weight model in one harness, that harness can now be a training target. The published figures say a 27B model gets more correct answers in fewer steps after adaptation. Before relying on the efficiency figures, read the trajectories. The authors found the model writing reasoning-like text in a mode meant to skip reasoning, so part of the saving may be reasoning that moved rather than reasoning that disappeared.

---

## Also this cycle

- **[SpecBench: A tool to evaluate AI-enabled software development workflows](https://www.thoughtworks.com/insights/blog/generative-ai/specbench-tool-evaluate-ai-enabled-software-development-workflows)** · Yixiong Jiang, Chen Chen and Yuhua Li, Thoughtworks, November 13, 2026 · In one sample run per workflow under the same model label, a spec, plan and decompose workflow raised mean behavioral correctness from 92.25% to 95.00% over direct implementation. It also roughly doubled time (83.74 to 163.93 minutes) and used about four times the tokens (450K to 1,782K). The authors call this descriptive evidence from limited cases, and no harness or data is linked.
