Introduction
The Cairn AI agent discovered a zero-day vulnerability in the medical imaging tool MONAI Label, which attackers exploited to gain the highest level of access to a server at an East Asian university. Today is November 20, 2026, and I'm Muyan. In this episode, we'll also look at data handling risks posed by AI service providers and why more red team testing doesn't necessarily mean more complete protection. We'll cover what to check in a moment.
Today's AI Security Highlights
1. Cairn Discovers MONAI Label Zero-Day Vulnerability, Server Compromised
- What happened: On November 20, 2026, BleepingComputer reported on a case disclosed by Amazon in which attackers used the Cairn AI agent, an AI system capable of carrying out tasks, to discover and exploit a zero-day vulnerability in MONAI Label. A zero-day vulnerability is a vulnerability for which defenders have no patch available before a fix is released. In this attack, the attackers gained the highest level of access to a server at an East Asian university. The MONAI Label maintainers have provided a patch.
- Who is affected: Organizations using the MONAI Label medical imaging tool should take note, especially by identifying the servers on which it is deployed. The source material does not specify affected versions, so no particular version can be ruled out on that basis.
- Recommended actions: Identify MONAI Label deployments, confirm and apply the patch provided by the maintainers, and check servers for unusual access and privilege activity.
- Source: BleepingComputer
2. Dragos Finds Data Leaks and Privacy Settings Issues at AI Service Providers
- What happened: In research published on November 19, 2026, Dragos found that data had been leaked from at least three AI service providers. The research also found that some services failed to honor privacy settings or data retention policies.
- Who is affected: Organizations that send enterprise data to AI services for processing should pay attention to how providers actually handle and retain that data. If settings or policies are not honored, organizations may lose control over how their data is used.
- Recommended actions: Inventory the data sent to AI services, verify providers' privacy settings and data retention practices, and consider using decoy credentials designed to detect leaks.
- Source: Dragos
3. AI Agent Red Team Testing: More Tests Don't Guarantee Complete Protection
- What happened: On November 20, 2026, Help Net Security reported that researchers found AI agents retrained on a large number of prompt injection examples became more susceptible to hijacking in some tests. Prompt injection involves hiding instructions in content an agent processes in an attempt to change its intended behavior.
- Who is affected: Enterprises that deploy AI agents or use red team testing to assess agent security should take note. Test counts alone may not show whether defenses cover different attack types or whether agents can still perform their normal tasks.
- Recommended actions: Review the attack types covered by red team tests, resistance to repeated attacks, and agents' ability to perform normal tasks.
- Source: Help Net Security
4. US Government Accountability Office Warns That AI Affects Both Attack and Defense
- What happened: In a report published on November 20, 2026, the US Government Accountability Office said AI could create new cybersecurity threats and amplify the impact of traditional cyberattacks. The report also noted that AI can be used for cyber defense.
- Who is affected: Government agencies and critical infrastructure operators need to assess AI's impact on both attacks and defenses.
- Recommended actions: Incorporate AI-related attack and defense risks into existing risk assessments, and review the relevant defenses and response procedures.
- Source: US Government Accountability Office
5. Zscaler Partners With Anthropic and Mistral to Protect AI Applications
- What happened: On November 20, 2026, Zscaler announced partnerships with Anthropic and Mistral to help enterprises protect AI applications and agents. Zscaler's endpoint security product integrates Claude models to analyze attack paths and is expected to enter a limited preview.
- Who is affected: Enterprises evaluating AI applications, agents, and endpoint security tools can watch for further details about product capabilities and the preview. The source material does not provide a launch date for the limited preview.
- Recommended actions: When evaluating the product, review how it handles data, the scope of its permissions, and how its analysis fits into existing security workflows.
- Source: Zscaler
Closing
The key takeaway today is that AI agents have already appeared in real-world vulnerability attacks, and using AI services also requires managing data risks. Start today by checking whether MONAI Label is deployed and confirming its patch status, then review which AI services receive your organization's data. I'm Muyan. See you tomorrow.


























Comments