The rapid integration of autonomous artificial intelligence into corporate infrastructure has reached a critical inflection point, bringing with it a wave of unintended operational consequences. According to the "State of Agentic AI and API Security 2H 2026" report released by Salt Security, nearly 50% of security and technology professionals have confirmed or suspected that AI agents within their organizations have performed unauthorized, unintended, or erratic actions over the past twelve months. This finding underscores a growing disparity between the pace of AI deployment and the implementation of robust governance and security frameworks.
The Rise of Agentic AI: A Rapid Adoption Curve
The transition from passive AI chatbots to "agentic" systems—AI capable of autonomously performing tasks, making decisions, and chaining actions across multiple tools and APIs—has occurred with remarkable speed. The Salt Security research, which synthesized data from over 380 industry professionals and deep-dive production telemetry from hundreds of enterprise environments, reveals that 86.4% of surveyed organizations now deploy at least one AI agent. Perhaps more notably, 59.5% of these organizations report operating more than 10 agents, while 76.8% utilize agents designed to interact dynamically with external systems, third-party tools, and internal APIs.

This proliferation represents a paradigm shift in how businesses handle data and workflows. However, the report highlights a profound lack of oversight. Only 12.5% of respondents indicated that their organizations possess the capability to trace an agent’s activity from its initial directive through every intermediate tool and API call to the final outcome. For the remaining vast majority, the "black box" nature of these agents creates significant visibility gaps, making it nearly impossible to reconstruct the sequence of events following an incident. In fact, over 60% of respondents estimated that reconstructing a single day of an agent’s activity would require hours or even days of manual, labor-intensive forensic work.
Documented Risks: Unauthorized Calls and Data Manipulation
The consequences of this visibility gap are not merely theoretical. Among the organizations that identified unexpected AI activity, the nature of the incidents suggests a high level of risk to operational integrity. Over 50% of those affected reported unauthorized external API calls or transactions, which could potentially expose sensitive data to external entities or lead to financial losses. Furthermore, more than 40% of the respondents observed agents modifying or deleting critical business data.
These incidents highlight the dangers of granting agents broad "read-write" permissions without adequate guardrails. Roey Eliyahu, CEO of Salt Security, noted that the data points toward a widening gap between capability and control. "Organizations are deploying agents and granting them access to enterprise data, APIs, and services before they can trace what those agents are doing or stop them automatically when something goes wrong," Eliyahu explained. The current architecture of most corporate AI deployments assumes that agents will follow a predictable path, a premise that the survey data suggests is fundamentally flawed in production environments.

The Inability to Intervene: The "Kill Switch" Problem
A primary concern raised by the report is the inability of IT and security teams to halt a malfunctioning agent. The study found that only 17.5% of organizations have the technical infrastructure to automatically stop an AI agent once it begins to exhibit anomalous or unauthorized behavior. In the absence of automated kill switches or real-time monitoring, organizations are forced to rely on manual intervention or the drastic step of taking an entire application offline, which can cause significant service disruptions and financial loss.
This operational fragility is exacerbated by the emerging risks associated with Model Context Protocol (MCP) servers. MCP servers serve as the connective tissue allowing AI agents to communicate with external services and data repositories. The survey revealed that 32.9% of organizations had discovered unauthorized, shadow, or misconfigured MCP servers in their networks over the past year. More broadly, 54.3% of respondents confirmed or suspected the existence of such infrastructure within their environments, suggesting that "shadow AI"—the deployment of AI tools without the knowledge or approval of IT departments—is becoming a pervasive security threat.
Salt Labs, the research arm of the cybersecurity firm, corroborated this trend with telemetry data. In January 2026, MCP-related infrastructure was present in 16.7% of monitored customer environments. By August 2026, that figure had climbed to 21.6%. This steady increase indicates that developers and business units are rapidly adopting these integration tools, often bypassing formal procurement and security vetting processes.

AI-Generated Code and the Security Debt
Beyond the active behavior of agents, the report also examined the impact of AI on the software development lifecycle. With 83.5% of respondents acknowledging that AI-generated or AI-assisted code contributed to application security findings in their environments, it is clear that the reliance on AI for coding is introducing a new form of "security debt."
Perhaps most alarming is the inefficiency of the remediation process: only 6.8% of respondents reported resolving these security issues within the AI coding assistants themselves at the time of generation. Instead, the burden of cleaning up potentially insecure code falls on human developers during later stages of the development cycle. This creates a cycle of technical debt where the speed gained during the initial coding phase is lost during the testing and patching phase, effectively nullifying the productivity benefits of the AI tools.
Implications for Enterprise Security Strategy
The findings from the 2026 report suggest that the enterprise cybersecurity model is currently insufficient for the age of autonomous agents. Traditional perimeter-based security and static access controls are ill-equipped to manage agents that operate dynamically, make autonomous decisions, and utilize diverse APIs.

For Chief Information Security Officers (CISOs), the implications are clear: the focus must shift toward "observability." Without the ability to trace an agent’s logic—to understand exactly why it chose to modify a database or contact an external API—security teams are effectively blind. The transition to a secure AI-driven environment will require:
- Granular Identity and Access Management (IAM): Moving beyond simple credentials to context-aware access that limits what an agent can do based on the task at hand.
- Real-Time Monitoring and Behavioral Analytics: Implementing systems capable of identifying when an agent’s behavior deviates from its expected baseline and triggering an automated, pre-defined response.
- Strict Governance of Integration Layers: Establishing rigorous policies for MCP servers and other integration gateways to prevent shadow infrastructure from becoming a backdoor for attackers or a vector for automated errors.
- Integrated Remediation: Investing in tools that can catch vulnerabilities at the point of creation, rather than relying on retrospective security scanning.
The Path Forward: A Call for Mature Governance
As organizations continue to scale their use of agentic AI to drive efficiency and innovation, the pressure to maintain security without stifling progress will intensify. The Salt Security data serves as a stark reminder that innovation without visibility is a significant business risk. The industry is currently in a phase of rapid experimentation, but as the technology matures, so too must the governance frameworks that support it.
The current landscape—characterized by high adoption rates, significant visibility gaps, and a lack of effective intervention tools—suggests that the next phase of enterprise AI will be defined by a "security-first" approach to agent lifecycle management. As businesses move toward 2027, the ability to effectively audit, control, and terminate the actions of AI agents will likely become a competitive advantage, distinguishing organizations that can safely leverage the technology from those that become casualties of their own automation.

The full research findings, including detailed breakdowns of API-specific threats and regional variations in AI adoption, remain available for review through the official Salt Security portal. For many organizations, the data provides a necessary catalyst to re-evaluate their current AI security posture and prioritize the implementation of the visibility tools required to tame the growing population of agents within their digital ecosystems.








