AI agent access moves to real-time, task-based controls

In July 2026, Secure Agentics launched Adrian as an open-source runtime authorization layer for agents, pitched around stopping actions before execution rather than auditing them afterward. The company described the model as “gate before, don’t log after,” with Adrian reading an agent’s reasoning, checking each tool call against a defined remit, and blocking out-of-bounds actions in the moment, while also hardening the checker itself in zero-trust fashion so the guardrail could not simply become the next attack surface.

By late August, that runtime-access posture had moved from toolkit pattern to shipping enterprise product: “Britive’s August 24, 2026 launch release introduced ARC as an immediately available part of its platform,” designed to “give AI agents temporary, task-scoped privilege” and “remove access when the task ends or conditions change.” The same late-summer window also saw formalization beyond enterprise software: “The Ethereum standard ERC-8196, named ‘AI Agent Auth Wallet’, has completed its formal review process, and its technical specification has now been released for developers to implement,” explicitly aiming to “replace broad credential delegation” with policy-scoped control. Days later, ARISE framed “Agentic Runtime Identity Security Enforcement” as a distinct discipline for governing agent actions while executing, underscoring that by early September the market was converging on runtime, adaptive authorization as a category, not an isolated product feature.

The mechanism is operational, not theoretical: once agents act at machine speed, static scopes and human review stop being meaningful control points. In the Anthropic espionage campaign, “The AI made thousands of requests per second… Human operators intervened at 4-6 critical decision points… For 80-90% of tactical operations, no human touched the keyboard,” so the security boundary shifts to runtime authorization that is task-bound, intent-aware, and continuously re-checked rather than granted once and trusted thereafter.

That imperative is colliding with enterprise scale: Gartner predicts that by the end of 2026, 40% of enterprise applications will feature embedded AI agents, while “92% of organizations lack full visibility into their AI identities,” 86% do not enforce access policies for them, and only 5% feel confident containing a compromised agent. With non-human identities already outnumbering humans 45:1, 97% carrying excessive privileges, and 71% of credentials not rotated on schedule, the market is turning to centralized policy engines, short-lived transaction tokens, distinct agent identities, and in-line approval flows to reconcile autonomy with auditability and compliance. UK AISI reported analyzing over 1,000 publicly available Model Context Protocol servers and observed a sharp increase in new servers from December 2024 to July 2025, with the sharpest jump coming from June.