Agentic security: Enterprises rely on OpenAI most to secure AI agents, and those that do are nearly three times as likely to say attackers are ahead

What happened

VentureBeat Intelligence conducted a survey of enterprises with 100+ employees to assess how they secure AI agents. The findings reveal that OpenAI's guardrails are the most commonly used primary security layer, with 40% of organizations relying on them. The survey also found that enterprises using OpenAI are more likely to perceive attackers as having an advantage. Additionally, 59% of enterprises reported security incidents or near-misses involving AI agents, and 62% allow agents to share credentials, which poses significant security risks.

Why it matters

The findings highlight the critical need for robust security measures in AI agent deployment. The reliance on OpenAI's guardrails, despite the perception of attackers being ahead, underscores the challenges enterprises face in securing AI systems. The high incidence of security incidents and the widespread use of shared credentials indicate that current security practices may not be sufficient to protect against evolving threats.

• OpenAI's guardrails are the most widely used security layer for AI agents, but enterprises using them are more likely to feel attackers are ahead.
• A significant portion of enterprises experience security incidents with AI agents, emphasizing the need for stronger security measures.
• Shared credentials among AI agents pose a major security risk, making it difficult to trace and contain breaches.
• Sandboxing and runtime enforcement of permissions are underutilized, indicating gaps in enterprise security practices.
• Enterprises should prioritize credential management and isolation strategies to enhance AI agent security.

The bigger picture

The survey results indicate that enterprises are increasingly relying on major cloud providers for AI agent security, but many are still vulnerable due to shared credentials and insufficient isolation measures. This highlights the need for improved security strategies, including better credential management and the adoption of sandboxing techniques to mitigate risks associated with AI agents.

The survey by VentureBeat Intelligence reveals that while OpenAI's guardrails are the primary security layer for many enterprises, there are significant security challenges, including the perception of attackers being ahead and the prevalence of shared credentials. These findings underscore the importance of adopting more robust security measures to protect AI agents from potential threats.

Technical details

• OpenAI's guardrails and moderation tools
• Microsoft Azure's Purview data-protection tools
• AWS's Bedrock Guardrails
• Google Cloud's Model Armor and Vertex AI controls
• Anthropic's Claude Managed Agents controls
• Sandboxing for isolating high-risk agents
• Runtime enforcement of agent permissions
• Shared API keys and human credentials for agent access

FAQ