---
title: 'In the News: November 19, 2026 (Extra 3)'
description: 'Prime Intellect says a hosted model was swapped for a refusal-stripped copy for 53 hours while its capability benchmarks did not move.'
canonical_url: 'https://darkfactory.dev/news/2026-11-19-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-11-19-extra-3.md'
collection: news
date_published: '2026-11-19T23:05:00-05:00'
date_modified: '2026-11-19T23:05:00-05:00'
---

# In the News: November 19, 2026 (Extra 3)


Prime Intellect has disclosed that its hosted DeepSeek V4.1 Flash endpoint served a tampered model for two days this month. Its own tests on the tampered weights found capability scores unchanged and refusals almost gone. Codex and Gemini CLI also shipped releases that change approval and command-permission behaviour.

## 1. Prime Intellect: a hosted model was replaced with a refusal-stripped copy for 53 hours

**[Notice regarding deepseek/deepseek-v4.1-flash, 8 to 10 November 2026](https://www.primeintellect.ai/security/advisories/deepseek-v4-1-flash-model-substitution)** · Prime Intellect, security advisory PI-SA-2026-001 · November 19, 2026

From 01:24 UTC on November 8 to 06:35 UTC on November 10, "an unauthorised party replaced a deployment behind Prime Intellect's deepseek/deepseek-v4.1-flash endpoint with an abliterated copy." The advisory defines that as a copy modified to reduce or remove refusal behaviour, built from publicly available weights. Requests served by that deployment were answered by the modified model. The chat template, tokenizer and serving configuration were not changed.

Prime Intellect tested the modified weights. Capability benchmarks (GPQA Diamond and τ-bench) were unaffected. Refusals fell from 98 of 100 prompts to 1 on JailbreakBench, from 304 of 313 to 2 on StrongREJECT, and from 198 of 240 to none on HarmBench. The company says it found no evidence that prompts or completions were recorded or left the serving cluster, that no other models or infrastructure were reached, and that no keys or customer credentials were affected. Affected customers have been contacted, and the investigation is ongoing. These findings are the company's own, and the advisory does not say how the attacker gained access or how the swap was found.

**Why it matters:** The model behind a hosted endpoint changed for two days, and a check built on capability benchmarks like these would not have shown it. If your agents run on a third-party open-weights endpoint and you count on the model's refusals as one of your safety layers, that layer needs a test of its own.

## 2. Codex 0.171.0 makes GPT-6.2 Sol the default catalog choice and adds remote noninteractive runs

**[Codex rust-v0.171.0](https://github.com/openai/codex/releases/tag/rust-v0.171.0)** · OpenAI, GitHub release notes · November 20, 2026, 00:44 UTC

The release adds GPT-6.2 Sol and Astra to the model catalogs, "with Sol as the default catalog selection." A new `codex exec --remote <ADDR>` starts or resumes noninteractive sessions on an existing app server. Two fixes concern approvals: late completion events no longer cancel approvals for a newer turn, and permission profiles now stay intact when the working directory changes. The TUI workspace, which shows subagent transcripts beside the main conversation, is now on by default.

**Why it matters:** If your scripted Codex runs take the catalog default instead of pinning a model, confirm which model they get after this upgrade. Teams that drive Codex from another process can now point noninteractive runs at an app server that is already running.

---

## Also this cycle

- **[Gemini CLI v0.70.0](https://github.com/google-gemini/gemini-cli/releases/tag/v0.70.0)** · Google, GitHub release notes · Among the merged fixes: system MCP servers with insecure command permissions are now skipped, shell commands with unresolved expansion keep the `ASK_USER` decision, and `commandPrefix` and `commandRegex` rules are matched against the top-level command argument.
