Introduction
Microsoft Agent Framework packages on PyPI have reportedly been hijacked, and teams using the affected packages should take notice. Today is November 17, 2026. I'm Mu Yan, and in this episode we'll also look at developments in AI agent risk management, including what to check during execution beyond taking inventory of permissions.
Today's AI Security Highlights
1. Microsoft Agent Framework packages reportedly hijacked
- What happened: According to OX Security, Microsoft Agent Framework packages on PyPI, the Python package publishing platform, were hijacked. The available source evidence is not yet sufficient to confirm the affected versions, disclosure date, or remediation status, nor can it verify whether credentials were stolen.
- Impact: Development teams using the relevant packages should check their environments, especially which hosts have the packages installed and which cloud accounts and secrets those hosts can access. Finding that a package was installed is not enough to conclude that credentials were exposed; logs must also be investigated.
- Defensive recommendations: Start by taking inventory of the packages and related hosts. Check access logs and unusual activity, then address potentially affected credentials based on the investigation and verify package version information with trusted sources.
- Source: OX Security
2. Wiz draws attention to hidden risks in AI agents
- What happened: Wiz published material about its AI Security Agent focused on finding hidden risks. The currently available source evidence is insufficient to confirm the product's public preview status, release date, or specific features.
- Impact: When organizations adopt AI agents, programs that can carry out tasks, they need to know what data the agents can access, which accounts they use, and which systems they can operate. Different teams may create agents, making it difficult for administrators to gain a complete picture of permissions without a shared inventory.
- Defensive recommendations: Create an inventory of AI agents that records their owners, purposes, and accessible systems and data. Regularly verify that their permissions match their work requirements.
- Source: Wiz
3. Pretxt unveils AI Supervision tool and secures seed funding
- What happened: SecurityWeek reported that Pretxt unveiled an AI Supervision tool and secured $5 million in seed funding. The available source evidence is not yet sufficient to confirm the tool's specific features or release date.
- Impact: Organizations that use AI agents to perform work or operate systems should consider logging, incident response, and assigned responsibilities during execution, in addition to reviewing permissions before deployment. These are general management recommendations and do not establish what capabilities Pretxt's tool provides.
- Defensive recommendations: Make sure every critical agent has an owner and activity logs. Decide in advance who will assess anomalies, how the agent will be disabled, and how incidents will be reported.
- Source: SecurityWeek
4. Security vendors expand AI agent protection, offering industry context
- What happened: PChome's November 7, 2026 security weekly roundup covered developments from several vendors, including SentinelOne's integration with Claude Cowork and Sysdig AI Defense's agent security and governance features on macOS and Windows. The report also listed CrowdStrike's capabilities for detecting and containing anomalous AI agent behavior.
- Impact: This earlier industry coverage offers organizations using AI agents a reference for potential protection strategies. When evaluating tools, operations teams should establish which devices and agents are actually covered and who will respond when an anomaly is detected.
- Defensive recommendations: Check whether existing monitoring and disablement procedures cover critical AI agents, and run exercises to confirm that owners can access the necessary logs and complete the response.
- Source: PChome (Sun Media)
Closing
The key takeaway today is that managing AI agents means looking at package sources, accessible credentials, and activity during execution. Today, ask your team to check whether it uses packages related to Microsoft Agent Framework, identify the hosts where they are installed, the secrets they can access, and the relevant access logs, then take action based on the investigation. I'm Mu Yan. See you tomorrow.


























Comments