The security of AI agents is a pressing concern for enterprises, with a majority already experiencing security events or near-misses. Despite this, confidence in agent security remains low, with 30% of enterprises believing AI-armed attackers are ahead of their defenses. The survey reveals a significant gap in agent containment, with only 18% of enterprises isolating their highest-risk AI agents, compared to 65% enforcing scoped permissions at runtime. This containment gap is a critical issue, as it allows for the propagation of a single control failure across systems. The majority of enterprises (53%) have agentic AI systems in production, and 53% have already had an agent security event, with 19% confirming an incident and 38% having identified a near-miss. The near-misses outnumber confirmed incidents, indicating that enterprises are catching problems but close to the edge. The survey also highlights a lack of confidence in agent security, with 63% of enterprises reporting credential sharing across their agent fleets. This sharing of credentials contributes to a growing lack of confidence in agentic security, as it makes it difficult to establish clean attribution and least-privilege access. The security stack is still heavily reliant on hyperscaler or model provider-native controls, with 92% of enterprises naming a primary security layer as a provider-native offering. This reliance on borrowed controls is a concern, as it may not provide the necessary level of security for AI agents. The survey also reveals a disconnect between satisfaction and urgency, with enterprises rating their current agent security tooling highly (4.29 out of 5) while planning to replace it within 12 months. This indicates that enterprises are more comfortable with the convenience and low friction of provider-native controls rather than confident in their demonstrated containment. The budget allocation for agent security is also a lagging but responsive indicator, with a majority of enterprises spending less than 10% of their security budget on agent security. However, the enterprises spending above a tenth are more likely to build scoped identity and isolation controls, which could lead to a narrowing of the containment gap. The arms race between AI-enabled defenses and AI-enabled attackers has tilted, with 63% of enterprises rating the balance as even or worse. This pessimism is driven by experience, with enterprises that have had a confirmed incident or near-miss being more likely to say attackers are ahead. The survey also highlights a lack of interest in agent-identity products and runtime sandboxing tooling, which are critical controls for addressing the containment gap. In conclusion, the security of AI agents is a complex issue, with a need for a comprehensive approach that includes both containment and prevention. Enterprises must address the containment gap to ensure the security of their AI agents and protect against the growing threat of AI-armed attackers.
