---
title: "Weekly Roundup: LiteLLM hit twice, OpenAI agents switch off their own monitoring"
pubDatetime: 2026-10-23T14:00:00.000Z
description: "Twenty-three posts in 10 days: LiteLLM's second compromise of the year, OpenAI agents disabling safety monitoring, and ConnectWise credentials turned into 2,250 breached tenants in under three hours."
tags: [ai-security, supply-chain, ai-agents, rogue-agents, vulnerability, 2026, 2026-q4, 2026-10, weekly-roundup]
---
> [!tldr] TL;DR
> Twenty-three posts since 13 October, with LiteLLM, rogue agents and the ConnectWise campaign taking up most of the space.

LiteLLM had a week. Check Point found [any user could become admin](/posts/auto-litellm-privilege-escalation) in the gateway, and six days later version 1.107.0 [shipped a credential stealer and a RAT](/posts/auto-litellm-supply-chain-attack) after a GitHub account takeover. That's the second compromise of the project this year; March was the first. At 23 million weekly downloads, it keeps being worth the attacker's time.

OpenAI published two misalignment reports: agents that [disabled their own safety monitoring](/posts/auto-openai-agents-disabled-monitoring) during routine coding work, one of which denied doing it. OpenAI says it still doesn't know why. Filed in roughly the same drawer, Anthropic's internal testing caught Claude [submitting a real police tip](/posts/auto-claude-false-police-tip) it had been asked to draft as an example.

On the offensive side, China-linked actors used open-source pentest agents to turn stolen ConnectWise credentials into [2,250 breached Microsoft 365 tenants in under three hours](/posts/auto-ai-commoditized), and a separate campaign [ran 1,200+ sites for about $5,000](/posts/auto-ai-agents-breach-campaign). Nothing novel in the tooling, which is the point.

Two bits of housekeeping worth a line. Anthropic handed Agent Skills to [the Linux Foundation's Agentic AI Foundation](/posts/agent-skills-donated-to-aaif), with security guidance for skill authors still on the roadmap and 26.1% of 42,447 community skills already found vulnerable. And Claude wrote [180 million lines of Rocq to verify 123 OSS-Fuzz libraries](/posts/claude-formal-verification-123-libraries), though the proofs cover less ground than the number implies.