OpenAI leads enterprise AI agent security, while most teams still share credentials

A VentureBeat Intelligence survey says 40% of organisations naming a primary security layer for AI agents rely on OpenAI’s guardrails, up from 21% in July. But the same survey points to a basic gap beneath the vendor rankings: 62% of organisations running agents in production said some or all of those agents share credentials.

Watch Desk analysis

What happened

The survey, conducted in August among people at organisations with 100 or more employees, found Microsoft Azure was the primary agent security layer for 22% of respondents who named one. Together, OpenAI, Microsoft Azure, Google Cloud, Anthropic and AWS accounted for 99% of those answers.

VentureBeat Intelligence also found that 59% of respondents whose organisations run agents in production or pilot reported an agent-related security incident or near-miss in the previous year; 30% reported a confirmed incident. Among organisations running agents in production, 38% said every agent had its own identity, while the rest reported at least some shared or human credentials. The report’s survey findings on agent security include respondents’ accounts, not an independent measurement of each organisation’s security.

Key findings

• OpenAI was the most common primary layer
  It was named by 40% of respondents who chose a primary layer, compared with 22% for Microsoft Azure.
• Incidents and near-misses were widely reported
  59% of respondents at organisations running agents said one had caused an incident or near-miss in the past year.
• Shared credentials remain common
  62% of production-agent organisations said some or all agents use shared or human credentials.
• Sandboxing was rarely the stated priority
  Only 9% said isolating high-risk agents in a sandbox best described their security programme.

Why it matters

A provider’s built-in controls may be the first line of defence for many organisations, but that does not settle the question of who can trace an agent’s actions or switch off one misbehaving system without disrupting others. Shared credentials can make both harder, particularly when an agent inherits a person’s access.

The survey also found that 43% of respondents relying primarily on OpenAI said attackers using AI were ahead of their defences, compared with 16% relying on another primary layer. That is a difference in respondents’ perceptions, not proof that OpenAI’s controls cause weaker security or that those organisations are objectively more exposed. The report says incident rates between the groups were too close to call.

Our read

The headline is OpenAI’s lead. The more practical lesson is to ask what that lead is being asked to protect: named identities, tightly scoped permissions and a way to isolate risky agents are less glamorous than a vendor ranking, but rather more useful when something goes wrong. A security layer is not a security strategy by itself.

What to watch

• Whether more organisations give every production agent its own managed identity.
• Whether reported use of sandboxing and other isolation measures rises.
• Whether future survey rounds find changes in incidents, credential practices or confidence in defences.

Discussion spark: Should organisations prioritise a provider’s built-in AI agent guardrails, or require independent controls such as per-agent identities and sandboxing as well?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.