Runtime governance rises as AI agents scale

By early 2026, autonomous agents were no longer experimental. VentureBeat said the “OpenClaw moment” was the first time such systems had “escaped the lab” into the general workforce, while the same February signal set also showed the industry moving from single agents to “agent teams”; days later, Resilient Cyber pointed to Boris Cherny’s viral example of 259 pull requests in 30 days, all AI-written, as evidence of a “fundamental shift” as enterprise and developer adoption accelerated.

That adoption quickly became an enterprise integration story, not a pilot story. In March, Ian Webster said virtually every CIO and CTO his firm spoke with had “some sort of agentic thing being built,” with roadmaps to connect agents to Salesforce and other internal systems, adding, “I definitely think I’m on board with 2026 being the year of the agent”; by September, Akamai warned of workplace agentic-AI blind spots, reporting that more than 40% of enterprise users had installed AI browser extensions and 25% of those extensions changed permissions within 12 months.

The scale problem became undeniable as enterprise AI spread faster than governance could mature. SiliconANGLE theCUBE reported that a September survey of 625 AI professionals across 13 industries, nine professions, and three countries found a trust index of “about 2.4 out of five,” with only “29%” saying they had enterprise-wide standardized frameworks in place; just “49% of AI professionals and end users actually trust the outcomes coming from AI agents,” a sign that human-era controls were not keeping pace with autonomous system access and decision-making.

At the same time, the attack surface was expanding at machine speed, not governance speed. CyberWire Daily said enterprises tracked 800 AI applications in December and 2,800 by May—“250% growth in just five months”—while “over half of enterprise employees now use generative AI apps daily” and “up to 30% of what they send contains sensitive data,” helping explain why the same SiliconANGLE theCUBE survey found the investment index rising to 3.8, with 73% planning significant or strategic trust-and-governance spending and 72% prioritizing data assurance and protection as the baseline control.

The governance failure is structural: enterprises are trying to manage autonomous agents with identity systems built for people. On SiliconANGLE theCUBE, Jack Hirsch said “over 90% of organizations have AI agents already deployed and generally running amok” while only “about 10% have a control or governance strategy,” and he argued the two common patterns—static API keys and end-user OAuth grants—either create unmanaged non-human identities or shift authorization to employees, leaving CISOs unable to see which agents are requesting access to which enterprise resources.

That mismatch shows up in both incidents and readiness data: AI + a16z described a 2026 case where an agent returned other companies’ data because it used the wrong identity context, a plain authentication-and-authorization failure rather than simply “more AI.” Portnox/Wakefield found 78% of CISOs lack a formal strategy for AI identities, while NeuralTrust reported 73% are highly concerned but only 30% have mature safeguards; 46% remain reactive, 29% managed, 1 in 5 organizations has experienced at least one AI agent–related security breach, mainly driven by prompt injection and data exposure, 1 in 5 have already had an agent-related breach, 42% mainly monitor activity, 38% rely on access control, and 25% still have no AI-specific controls.

By early 2026, the market response had shifted from static approval gates to controls that sit inside the agent loop, because machine-speed execution was outrunning human review. Resilient Cyber captured the break point with examples of “users running agents in auto-approve mode for 45 minutes at a time,” “agents disclosing SSNs and destroying infrastructure,” while RockCyber Musings noted OWASP’s Top 10 for Agentic Applications, released in December 2025, tied agent-specific risks such as “memory poisoning, tool misuse, and privilege compromise” to authorization-layer failures—evidence that one-time access decisions were no longer enough once agents were acting continuously.

Vendors and enterprises then began converging on always-on governance layers: Palantir emphasized “runtime policy evaluation,” Operant launched “the first real-time agentic security solution,” and standards work accelerated with “NSS Labs, AWS, Microsoft, F5 Set New AI Security Standards With Two Foundational White Papers.” The urgency came partly because formal guidance would lag—“Best case from the current comment cycle: interim guidance in twelve months”—so Microsoft moved to operational tooling, with Satya Nadella saying, “Starting with observability… something called Agent 365,” extending Entra, Defender, and Purview into a unified runtime control plane.