April 18, 2025

⚡ Key takeaways

• DeepSeek's cyber attack forced temporary registration limits, revealing vulnerabilities in AI service providers
• As AI agents make autonomous decisions (travel booking, purchases, scheduling), they create new security risks beyond traditional data breaches
• Organizations are seeing unauthorized employee connections to public LLM services, creating significant security and compliance risks

‍

The recent cyber attack against DeepSeek, which temporarily forced the Chinese AI company to limit new registrations, provokes us to think about the ramifications of AI adoption from a security perspective.

While DeepSeek's attack appears to follow familiar patterns of credential compromise, it signals a broader concern about security in the age of AI agents. Traditional cyberattacks typically target data theft or service disruption. However, with AI agents increasingly acting autonomously on our behalf, the threat landscape is evolving in concerning ways.

"While attacks like the one against DeepSeek follow familiar patterns, AI agents introduce fundamentally new security challenges," says Matt Wolf, Co-founder and Chief AI Officer at Obsidian Security. "When an agent operates autonomously on your behalf, compromising just a few data points in its decision-making pipeline can have far-reaching consequences. Organizations need to think beyond traditional security models to protect not just their data, but the entire context that influences how their AI agents behave."

‍

The Enterprise Security Challenge 

‍

The DeepSeek incident highlights a critical challenge for enterprise security teams. As AI tools proliferate, organizations are seeing increasing attempts by employees to connect to public LLM services - often without a proper security review or vendor risk assessment. On average, employees in large enterprises may use 5-20+ AI tools. This creates significant security and compliance risks that need to be managed.

‍

The real security concern with AI agents goes beyond traditional data breaches. These agents operate on extensive data sets (10x more than human identities), including chat histories and user preferences, to make decisions autonomously. This creates new attack vectors where bad actors could potentially:

• Manipulate the contextual data that agents use to make decisions
• Poison the input sources that agents rely on for information
• Influence agent behavior without directly compromising the system

‍

For example, imagine an AI agent tasked with booking business travel. If an attacker can manipulate the agent's understanding of your preferences or company policies, it could make decisions that appear legitimate but serve malicious purposes.

‍

Policy Considerations 

Organizations need robust AI usage policies that address both cloud-based and local LLM deployments. Key considerations include:

• Vendor risk management processes for new AI services
• Data privacy requirements for AI interactions
• Monitoring and oversight of AI tool usage
• Clear processes for requesting and approving new AI tools

‍

Protecting Against Agent-Based Threats 

Organizations need to approach AI agent security with particular attention to:

1. Data Pipeline Security: Strictly control and audit all dynamic inputs that influence agent decision-making
2. Context Integrity: Ensure the integrity of historical data and user preferences that agents use
3. Output Validation: Implement robust monitoring systems for agent actions, especially for autonomous operations
4. Access Controls: Implement strict policies around which AI services employees can access and use

‍

Final Thoughts

The DeepSeek incident serves as a wake-up call for organizations implementing AI agents. While AI drives innovation by automating processes and enhancing efficiency, it also introduces new security challenges that require innovative protection approaches. As AI tools become part of your daily operations, you need visibility and control over how they're being used.

Obsidian helps you stay ahead by automatically discovering every AI service your teams access, blocking unauthorized tools in real-time, and protecting your sensitive data from exposure. With browser-level security controls, you can confidently embrace AI innovation while maintaining strong security guardrails - we've already helped customers block thousands of unauthorized AI access attempts, including many that traditional security tools missed.

The future of AI security isn't just about protecting data - it's about ensuring the integrity of the entire decision-making pipeline that powers our AI agents. As we continue to see rapid AI adoption across enterprises, staying ahead of these emerging threats means implementing both robust security measures and comprehensive policies that govern how AI tools are adopted and used within your organization.

‍