DAILY BRIEFING
AI "Mind Viruses" Can Spread Between Agents Through Persistent Prompt Files
5 stories2 min read
Today's threat landscape highlights how attackers leverage legitimate cloud services and modern software architectures to bypass traditional defenses. Threat actors are repurposing trusted platforms like Microsoft 365, Salesforce, and ServiceNow to hide command-and-control operations and conduct sustained data harvesting. Security teams operating enterprise SaaS environments should prioritize auditing API traffic and reviewing guest permissions first.
At the same time, emerging research into autonomous AI agents demonstrates novel propagation risks through persistent state files. Meanwhile, software supply chains remain a target for credential theft via typosquatted package repositories.
TWINLOOT Abuses SharePoint and Teams to Steal Credentials and Move Across Networks
Researchers disclosed TWINLOOT, a modular PyArmor-hardened Python framework that routes C2 traffic entirely through trusted Microsoft 365 infrastructure. It uses headless Microsoft Edge instances, Graph API file dead-drops on SharePoint, and WebRTC DataChannels over Microsoft Teams TURN servers to deliver malware.
One Attacker Has Scraped Both Salesforce and ServiceNow Portals Since 2025
A single piece of infrastructure has been pulling records out of Salesforce and ServiceNow customer portals across multiple industries for more than a year, according to research published this week by agent security platform Reco.
- Why it matters
- Organizations relying on Salesforce and ServiceNow customer portals face unauthorized automated scraping and bulk data exfiltration of corporate records.
16 Typosquatted RubyGems Packages Steal Browser Credentials and Crypto Wallets
A malicious campaign dubbed StubMaker published 16 typosquatted RubyGems packages targeting popular dependencies like i18n and activesupport. The clumsy typosquats deploy Windows-based information stealers designed to harvest browser credentials, crypto wallets, seed phrases, and Telegram data.
- Why it matters
- Ruby developers who accidentally mistype package names during dependency management risk infecting Windows endpoints with credential-stealing malware.
AI "Mind Viruses" Can Spread Between Agents Through Persistent Prompt Files
Researchers at Anthropic and EPFL demonstrated that self-propagating prompt payloads ('mind viruses') can spread between autonomous AI agents using editable system prompt files to carry state. While tested on coding agents and OpenClaw models, defensive prompt warnings effectively stopped propagation beyond one hop.
- Why it matters
- Organizations deploying autonomous AI agents that persist state through system prompts face risks of prompt injection payloads propagating across agent interactions.
Microsoft tests faster Windows File Explorer, new context menu
Microsoft has begun testing performance improvements for Windows File Explorer alongside a simplified, customizable context menu in Windows 11 preview builds released to Windows Insiders.
- Why it matters
- IT administrators and Windows 11 users will experience altered desktop navigation and potential changes in workflow context menu behavior upon full release.
- What to do
- Test Windows 11 preview builds in evaluation environments to assess UI and context menu updates.
Key takeaways
- Ontinue revealed TWINLOOT, a Python implant using Microsoft Graph API, SharePoint, and Teams for command-and-control traffic.
- Anthropic and EPFL demonstrated self-propagating prompt payloads across autonomous AI agents, though adding warning instructions mitigated spread.
- Reco identified the City Forum campaign using IP 158.220.87.79 to scrape Salesforce and ServiceNow portals since March 2025.
- Microsoft began testing File Explorer performance improvements and context menu options in Windows 11 preview builds.