A VentureBeat Pulse Research survey of 107 enterprises reveals that 54% have experienced an AI agent security incident or near-miss. Only 32% give each agent its own scoped identity, and most agents share credentials, widening the blast radius. The findings highlight a gap between agent autonomy and security controls.
SynthePulse Insight · AI deep reading
Version 2 · 1 source
VentureBeat Pulse Research shows that over half of enterprises have experienced AI agent security incidents or near-misses. The core issue lies in 69% of companies allowing agents to share credentials, while only 32% have implemented independent identity management. This has resulted in a sandbox isolation rate for high-risk agents in large enterprises as low as 20%, revealing significant structural gaps in defense systems.
According to VentureBeat Pulse Research's survey of 107 enterprises, the expansion of AI agent autonomy has far outpaced corporate security control capabilities. Data indicates that 54% of enterprises have encountered AI agent security incidents or near-misses in the past year, with 18% being confirmed security events and 36% being near-misses intercepted before causing substantial damage. Only 42% of enterprises reported no such events. This data reveals the pervasive vulnerability in current enterprise AI deployments: agents are granted real access to systems and data, while controls intended to limit their behavior lag significantly behind.
Notably, there is a paradoxical inverse correlation between incident rates and enterprise scale. As enterprise size increases, risk exposure grows, yet containment capabilities actually decline. Medium-sized enterprises (101-1,000 employees) had an incident or near-miss rate of 49%, whereas large enterprises (over 1,000 employees) saw this spike to 63%. Meanwhile, the sandbox isolation rate for high-risk agents in large enterprises dropped from 35% in medium-sized firms to 20%, and satisfaction with security toolkits fell from 4.36 to 3.97. This suggests that large enterprises, which run the most agents and have the most complex systems, are precisely the group possessing the fewest critical containment measures (i.e., isolation capabilities).
The core weakness revealed by the survey is the lack of management for 'non-human identities.' Only 32% of enterprises can assign independent, scoped managed identities to each AI agent. Conversely, 69% of enterprises engage in some form of credential sharing across their agent fleets. Specifically, 48% of enterprises stated that while some agents have independent identities, credential sharing still exists, and another 32% primarily rely on shared API keys or borrowed human and service account credentials to run agents. This hybrid state makes it difficult to enforce the principle of least privilege and prevents clear attribution of specific actions to individual agents during post-incident forensics.
The chaos in identity management is directly linked to higher security risks. Data shows that enterprises with credential sharing experienced incidents or near-misses at a rate of 63.5% (47/74), while those where all agents possessed independent scoped identities had a rate of only 40.9% (9/22). The difference is 23 percentage points. Although the sample size of enterprises fully implementing independent identities was small (22), and correlation does not imply causation, such a significant disparity in a single survey indicates that the lack of independent identities is a key structural factor leading to agent privilege escalation or expanded impact after compromise.
In terms of specific control measures, observation and enforcement are common, but critical isolation measures are severely lacking. Only 30% of enterprises implement sandbox isolation for their highest-risk agents to limit the potential radius of damage. This lack of isolation capability, combined with widespread credential sharing, means that if an agent is compromised or misconfigured, its attack surface will rapidly spread across the entire network.
Furthermore, enterprise security technology stacks show an over-reliance on cloud vendor native solutions. OpenAI's guardrails dominate with a 51% usage rate, while Google and Microsoft cloud controls and Anthropic's managed agent controls are also widely used. In contrast, the adoption rate of specialized security products designed specifically for agents is extremely low. Although this 'borrowed' stack currently enjoys high user satisfaction (average 4.2/5, based on 82 respondents), it reflects a lack of market maturity. Due to the absence of dedicated deep-defense tools, enterprises often rely only on basic protections from model providers, which proves inadequate when facing increasingly complex AI-driven attacks.
Despite severe security gaps, spending on security remains a small fraction of overall budgets. However, satisfaction scores remain high at 4.2/5, indicating a state of 'comfortable danger.' Enterprises seem satisfied with current native solutions, even though these solutions do not provide sufficient isolation and fine-grained identity control.
However, this satisfaction may be short-lived. Most enterprises plan to replace their existing security tools within one year. This implies that the current lightweight defense system based on cloud vendor natives cannot meet long-term needs. As AI agents play a more central role in enterprise operations, identity governance, fine-grained isolation, and the integration of specialized security products will become inevitable trends. The current survey sample is predominantly medium-sized enterprises (42% are 251-1,000 employees) and uses non-probability self-selection sampling; therefore, the results reflect the perspective of mid-sized organizations actively building agent security systems rather than the full picture of all large operators.
Data source is a single cross-sectional survey conducted by VentureBeat Pulse Research in June 2026, with a sample size of 107 enterprises, excluding those with fewer than 100 employees. The sample skews toward the mid-market (42% are 251-1,000 employees), uses non-probability sampling, and results should be viewed as directional signals rather than precise measurements. Satisfaction scores are based on 82 valid respondents.
The primary contradiction in AI agent security has shifted from 'whether to introduce agents' to 'how to control them.' The 54% incident rate warns enterprises that current credential-sharing models and weak isolation measures (only 30% sandboxed) are insufficient to handle risks. Independent identity management and dedicated isolation tools are key to closing security gaps. Although the current market still relies on cloud vendor native solutions, enterprises are planning to upgrade their toolchains within the next year.
Primary report
Primary source