An analysis of the corporate AI agent security market has revealed an alarming trend: despite the active deployment of such systems, the protection of their infrastructure remains highly fragmented. In a survey conducted in June 2026 among 107 companies with more than 100 employees, 18% of respondents reported a confirmed security incident, and another 36% reported a successfully prevented attack. Only 42% of participants had not encountered such threats, while 5% of organizations did not operate AI agents in a production environment at all.
The most vulnerable link turned out to be identity management practices. Only 32% of companies provided each agent with its own managed identity with limited permissions. Meanwhile, 48% of organizations implemented individual identifiers for only some agents, while the rest continued to use shared credentials. Moreover, 32% of respondents admitted that their agents still operate through shared API keys, employee accounts, or service accounts. As a result, it is estimated that 69% of surveyed companies use shared credentials in at least part of their infrastructure.
The correlation between credential separation and incident frequency is evident: among organizations using shared identifiers, 63.5% experienced an incident or prevented a threat. In comparison, in companies where each agent has a separate identity, this figure stands at 40.9%. Although the study authors emphasize that this does not prove a direct causal link, the data points to a clear vulnerability. However, the group with separate identifiers is small—only 22 participants.
Protection: Reliance on Vendors and Budget Constraints
82% of respondents cited built-in security tools from AI model providers as their primary line of defense (OpenAI — 51%, Google — 36%, Microsoft — 35%, Anthropic — 29%). The average satisfaction with current solutions is high—4.2 out of 5 points. However, 59% of organizations plan to implement new tools within a year, and 29% within the next three months. Notably, among companies that have already experienced an incident, the share of those wanting to upgrade protection in the near future reaches 42.1%, compared to 14% among those unaffected.
Spending on AI agent protection remains modest: 46% of companies allocate 6–10% of their cybersecurity budget, and 34% spend less than 5%. Only 24% allocate more than 10%.
Limitations and Conclusions
Although the survey covered key decision-makers (45% had the final say), it is skewed toward mid-sized businesses (67% are companies with 101–1000 employees). Nevertheless, the trend is clear: the pace of AI agent adoption significantly outpaces the development of practices for their isolation, access control, and permission distribution. The study does not assess the actual frequency of incidents in the corporate sector, but the signal is alarming.
My expert assessment: Relying solely on vendors' built-in tools is a risky strategy. As the survey showed, even with high satisfaction with existing solutions, nearly 60% of companies are ready to change them. The Zero Trust approach, proposed by Anthropic, with a separate identity for each agent and minimal access rights, is becoming not just a recommendation but a necessity. Ignoring this principle could lead to the next wave of attacks targeting vulnerable shared credentials.