Friday, November 20, 2026 Edition No. 244
Search
Daily News Fully automated · 4 min read

AI Security News: November 18, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. A Malicious Version of Microsoft’s agent-framework Package Was Published to PyPI With a Credential Stealer

Wiz reports that versions 1.26.1 and 1.26.2 of the agent-framework package (Microsoft Agent Framework) on PyPI contained credential-stealing code, published on November 17 and removed within roughly an hour. Version 1.26.1 installs a .pth file that runs on every Python interpreter start, even if the package is never imported, and 1.26.2 adds theft of GitHub Actions OIDC tokens. The malware targets environment variables, Azure CLI, npm, pip, git, Docker and Hugging Face credential files, and Azure managed-identity tokens, and sends the data to a webhook.site URL. The root cause is not confirmed: there is no matching release tag in the microsoft/agent-framework repository and 1.26.1 shipped as a wheel only. Wiz recommends pinning to 1.26.0, checking for the persistence files and rotating all secrets in affected environments. Source

2. Stormshield Details Trigram, a Campaign Impersonating About 50 AI and Blockchain Brands to Get Developers to Run an Infostealer

Stormshield published a technical report on Trigram, which it describes as a large-scale, industrialized campaign that impersonates about 50 brands in the blockchain and AI sectors to trick developers into executing an infostealer themselves. The Hacker News reports the campaign has targeted AI, blockchain and logistics sectors across Europe and the United States. Source

3. Apple Previews Finer-Grained macOS Permissions for Apps That Use AI Agents in macOS 27.3 Beta 2

Apple introduced controls that let Mac users give more granular access to data for apps, including those that use AI agents, according to The Hacker News. The changes are previewed in macOS 27.3 beta 2 and add transparency and options to two APIs, Device Control and Data Access, and Full Disk Access. Source

4. Simon Willison Describes an External Sandbox Pattern for Coding Agents With a Host Allow-List Proxy and Credential Replacement

Simon Willison wrote up an external sandbox pattern developed with Prime Radiant, where the coding agent runs outside the sandbox and any code it executes runs inside a separate one. The aim is to limit damage to the host from a mistake or a prompt injection attack when trying local models. The project includes three servers exposing the same localhost API, using v86 with Alpine Linux in WebAssembly, Apple Containers, and Microsoft’s Quicksand library with QEMU. An optional HTTP and HTTPS allow-list proxy limits which hosts sandboxed code can contact, and a credential replacement mechanism lets authenticated API calls work without giving the agent the secret. Pi, Claude Code or Codex can be pointed at any of them. Source

5. AgentShield Is an Apache-2.0 Local Firewall That Redacts Secrets and Blocks Destructive Shell Commands From Coding Agents

AgentShield, shown on Hacker News on November 18, is an open-source tool meant to stop coding agents such as Claude Code, Cursor and Windsurf from leaking .env secrets or SSH keys or running destructive shell commands. Its CLI blocks direct credential file reads and environment dumps, and a stream mode replaces secrets in output with placeholder tokens. The README says it has zero runtime dependencies, runs locally and air-gapped by default, and reports sub-20 microsecond P99 redaction in its own benchmark. The TypeScript repository is licensed Apache-2.0 and was created in October. Source