Report by Salt Security
State of Agentic AI and API Security 2H 2026
Key Findings
Only 12.5% of organizations can consistently trace a complete agent path across all agents, from initial instruction through tools and APIs to outcome.
Only 17.5% of organizations can automatically stop an agent that begins behaving unexpectedly in production.
32.9% of organizations have found an unauthorized, shadow, or misconfigured MCP server in the past 12 months.
Only 6.8% of organizations resolve AI-generated code issues inside the coding assistant at generation time.
49.8% of organizations confirm or suspect an AI agent took an action in the past 12 months that the organization did not intend, expect, or authorize.
86.4% of organizations report at least one AI agent operating.
59.5% of organizations run more than ten AI agents.
76.8% of organizations report agents that chain actions across other agents, tools, and APIs.
More than 60% of organizations would need hours or days of manual work to reconstruct what a specific agent did the previous day.
54.3% of organizations have confirmed or suspected an unauthorized, shadow, or misconfigured MCP server.
83.5% of organizations estimate AI-generated or AI-assisted code contributes to at least some of their application security findings.