Getty Images/iStockphoto
Securing AI meeting assistants: What IT leaders need to know
More companies are deploying AI agents to make their meetings more efficient. But these tools can also expose companies to some significant security risks.
AI meeting assistants have gained traction as workers use the tools to accelerate workflows and improve productivity. The meeting ends, and the AI assistant automatically sends the transcript and summary. That's a welcome departure from the days when any post-meeting information was shared as long videos and transcript files over email.
AI meeting assistants serve as moderators and co-hosts, handling meeting management and post-meeting services, such as recordings, note-taking, transcriptions and summaries. Assistants with agentic capabilities can understand the meeting context and provide suggestions to participants. This enables participants to focus on the meeting while the assistant takes notes, manages the agenda and tracks action items.
Agentic AI meeting assistants don't disappear after a meeting ends. They can coordinate schedules, set reminders for future meetings and follow up on action items.
Organizations can deploy two types of AI assistants for meetings:
1. Native UC platform meeting agents. The enterprise and its unified communications (UC) vendor jointly decide how to manage the data. RingCentral, Zoom, Cisco and Microsoft are some of the leading vendors in this area.
2. Third-party SaaS bots. Startups and small businesses might opt for standalone vendor products, such as Otter.ai, Fireflies.ai, Fathom AI, Read AI and Hedy AI, to simplify and streamline meetings.
AI assistant security risks to address
AI assistants for meetings offer many benefits, but they also raise security risks. Shadow AI, where workers use AI tools that aren't specifically sanctioned by their employers, is a growing issue. According to an IBM study, while 80% of American office workers use AI in their work, only 22% rely exclusively on company-provided tools.
Additional security risks include the following:
- Lack of enterprise-grade security. Bring-your-own-AI opens the door for employees to send sensitive corporate data to their personal AI accounts. IT and security teams not only lose control over sensitive data, but risk data misuse, financial losses and reputational damage.
- Governance concerns. AI assistants request to enter a meeting as an external participant through a calendar integration or OAuth permissions. The meeting data now resides in the vendor's cloud, which could potentially violate an organization's internal data policies. The vendor may use data collected by the assistant for internal purposes, such as AI training, which can further expose an organization's information.
- Consent laws. Enterprises that use third-party tools have to understand user consent laws and policies. Essentially, meeting attendees have to agree to be recorded. While some AI assistants for meeting tools seek consent, most vendors push these obligations onto the customer. How vendors collect and store this data is also under the microscope, especially biometric data that could be ripe for identity theft. A class action lawsuit filed late last year argues that Fireflies.ai violated the Illinois Biometric Privacy Act by collecting, possessing and retaining biometric voiceprints without consent.
Best practices for IT leaders
Enterprises must adopt a multi-pronged strategy to manage AI assistants, with clear data labels to prevent exposure. Common examples include digitally signed metadata, invisibly watermarked content and biometric authentication. Other components to consider:
- Establish AI policies. The first step is to establish policies, define rules and educate staff on AI. According to BlackFog research, 63% of employees believe it's acceptable to use AI if no corporate-approved list exists. IT leaders must ensure that only pre-approved AI bots are used for meetings, and that participants are made aware of their use beforehand. A common enterprise catalog and corporate AI accounts for all employees can ensure transparency in AI usage.
- Build a risk management model. The rapid rise of AI has prompted lawmakers worldwide to assess regulatory frameworks. However, AI regulations are a patchwork of state and federal rules. ISO/IEC 42001 is the first international standard governing AI management systems. NIST AI 600-1 is an AI risk management framework that safeguards enterprise deployment of generative AI (GenAI). Organizations can eliminate security vulnerabilities by using automated blocking tools when AI doesn't pass a vendor risk assessment.
- Own the agentic AI pipeline. Organizations can gain complete ownership of AI assistants for meetings by building their own agents internally. Workflow automation platforms, such as n8n or Zapier, can self-host the agent. The enterprise thus has 100% control through development and post-deployment. The pipeline can include transcription tools and use GenAI for summarization. Keep in mind, however, agentic AI deployment is complex and costly. Teams might lack the engineering expertise to build such AI tools.
- Control third-party tools. Cyber attackers target third-party AI applications, chatbots, meeting assistants and agentic pipelines to gain wider access to enterprise data. Attackers can manipulate meeting AI assistants with prompt injection attacks. Malicious prompts can make AI meeting agents silently record meetings and share them. Program agents to run on short-lived auto-rotating certificates rather than protocols like Model Context Protocol, A2A or AP2. Block third-party OAuth grants by default and reassess existing ones.
- Implement human-in-the-loop. If multiple AI agents join a meeting representing different human participants, for example, there is a risk that data or meeting summaries will be mismatched due to hallucinations. Manual review by humans to oversee AI workflows ensures data accuracy and enables them to intervene if something goes wrong.
Venus Kohli is an engineer turned technical content writer, having completed a degree in electronics and telecommunication at Mumbai University in 2019. Kohli writes for various tech and media companies on topics related to semiconductors, electronics, networking, programming, quantum physics and more.
