AI agents run wild as security controls lag behind

Venture Beat ↗

The gist

AI agents have exploded into the enterprise—outpacing security controls and leaving organizations dangerously exposed to identity chaos, overprivilege, and million-dollar breaches.

What to know

  • By fall 2026, over 90% of companies had AI agents running, but only 10% had any real control or governance in place.
  • Non-human identities now outnumber humans 144:1, with 97% dangerously overprivileged and most organizations blind to where and how LLMs are used.
  • Board-level risk is real: 1 in 5 firms has already suffered an AI agent-related breach, with many losses topping $1 million—and market leaders are scrambling to enforce agent-specific controls.

From Clawdbot to OpenClaw: The Tipping Point

A rapid leap from hobbyist bots to enterprise-wide agent teams unleashed powerful, unsupervised AI into core business systems before controls or certifications could catch up.

The surge was visible early. VentureBeat described an “OpenClaw moment” in February 2026, when autonomous agents moved from the hobby project “Clawdbot” in November 2025 to OpenClaw in late January 2026, and then into the general workforce with shell commands, file management, and Slack or WhatsApp access under persistent, root-level permissions. VentureBeat also linked early-2026 platform releases to a shift from single agents to “agent teams,” which accelerated deployment before governance had caught up. That same dynamic drove calls for explicit assurance mechanisms such as AIUC-1 certification for agents, because enterprises otherwise may be less ready to accept the consequences.

By the fall, rollout had clearly outrun control. On SiliconANGLE’s theCUBE, Okta’s 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 to discover, control, and govern AI agents. In August, N2K said agentic tools made failures no longer if, but when, with cascading errors 10 times faster and a larger blast radius, as non-human identities rose to roughly 109 for every human, up from about 80 just months earlier. That matched the September-through-November 2026 reports and surveys showing low policy enforcement and high breach rates.

Sources
Venture BeatSiliconANGLE theCUBEN2K Networks

Identity Sprawl Fuels Hidden Risks

Non-human identities have exploded in both number and privilege, with most agents retaining unchecked access and legacy security unable to detect or contain their reach.

The old model fails first on scale and identity. Spencer Thielman of Palo Alto Networks said his team cataloged 800 AI applications last December and 2,800 by May, a 250% jump in five months, while Entro Security found the non-human-to-human identity ratio had reached 144:1, up from 92:1; worse, security teams conducting audits discover 6-10x more identities than they tracked. That sprawl is not just inventory bloat: 97% of non-human identities have excessive privileges, 71% are not rotated on time, and 91% of former employee tokens remain active after offboarding, turning agents into durable, overprivileged footholds.

It also fails because agent risk sits in access paths and human workarounds that legacy controls were never built to govern. Joel de La Garza described a 2026 SaaS incident where an agent meant to return a user’s own company data instead returned other companies’ records when asked for “my data,” calling it an authentication and authorization failure; meanwhile Zscaler, analyzing billions of web records, found engineering drove 47% of AI traffic through coding copilots, and as RockCyber Musings put it, “Shadow AI isn’t driven by malice. It’s driven by friction” — if your secure, approved internal AI takes five minutes to load, requires a VPN, and hallucinates half the time, your executives are going to use ChatGPT, just as identity-centric compromise in Match Group’s January 30 breach — where parent of Tinder and Hinge confirmed a breach and threat actor group ShinyHunters claimed theft of 10 million records, with analysis stating the attack vector was “a social” — showed attackers can bypass technical hardening by exploiting the access layer itself.

Sources
CyberWire DailyRockCyber MusingsAI + a16zN2K NetworksRockCyber Musings

Governance Gaps Leave Blind Spots

Most organizations have written AI policies but lack the tools and visibility to enforce them, leaving critical agent activity and LLM usage invisible to security teams.

The control gap is no longer about whether enterprises have drafted AI principles; it is whether those principles can be enforced across the business. In a September survey of 625 AI professionals cited by SiliconANGLE theCUBE, Scott Hebner said only 29% had enterprise-wide standardized frameworks in place, with governance maturity at 2.8 and trust around 2.4 out of five; just 49% of professionals and end users trusted AI-agent outcomes, while 73% planned major trust-and-governance investments over the next 18 months, a sign that many programs remain more aspirational than operational.

That weakness is especially stark around visibility and identity, the two prerequisites for real oversight. A Portnox/Wakefield Research survey of 200 U.S. CISOs released in November 2025 found 78% lacked a formal strategy for handling AI identities in a zero-trust architecture, even though those identities include autonomous agents, bots and other machine actors, and the same share expected AI to create significant new security work; meanwhile, Harness’s State of AI-Native Application Security 2025 found 62% had no visibility into where LLMs were being used across their organizations, calling it the new enterprise blind spot.

Sources
SiliconANGLE theCUBEPR Newswire - Business TechnologyPR Newswire - Consumer Technology

Boardrooms Face Unquantifiable AI Fallout

C-suite anxiety is rising as breaches mount and losses escalate, yet few organizations have moved beyond reactive controls or achieved mature, proactive agent governance.

By late 2026, AI agent risk had clearly crossed into board territory because deployment was already broad while executive confidence was not. NeuralTrust’s The State of AI Agent Security 2026, cited by PR Newswire, found 73% of CISOs were very or critically concerned about AI agent risks, but only 30% had mature safeguards; the same study said most organizations were deploying agents faster than they could secure them, with 46% in a Reactive tier, 29% in Managed, and fewer than 10% reaching Proactive governance. It also found that 25% had no AI-specific controls at all, while 42% used activity monitoring, 38% relied on access control, 19% conducted adversarial testing, and 16% validated their AI supply chain.

The scale also shows up in damage and retrenchment, not just anxiety: PR Newswire reported that 1 in 5 organizations had already suffered an AI agent-related breach, and among those hit, 40% put losses at $1 million to $10 million while 13% expected more than $10 million. At the same time, Cority’s global survey found 95% of EHS+ teams were already using AI tools outside approved systems, and Thoropass said AI had become the top compliance and audit risk, reinforcing that the problem is already large enough to be a board-level risk.

Sources
PR Newswire - Business TechnologyBusiness WireGlobeNewswire - Industry News on Technology

Governance Drives the New AI Security Race

Demand for agent-specific controls is reshaping the security market, pushing vendors to deliver real-time discovery, least-privilege enforcement, and unified oversight as industry standards rapidly evolve.

The market is moving because governance has become the gating factor for enterprise agent adoption, not a downstream compliance add-on. Resilient Cyber, citing a Cloud Security Alliance survey, said “governance maturity is the strongest indicator of readiness,” with organizations that have formal governance twice as likely to adopt agentic AI and three times more likely to train staff; the same analysis said “the AI security landscape is bifurcating,” with genuine progress in agent identity standards, prompt-injection metrics, and policy-aware loops even as operational cracks widen, pulling buyers toward agent-specific control layers. It also concluded that “The rapid adoption of MCP has outpaced security considerations, and this governance framework attempts to address that gap,” citing the “VulnerableMCP Project” and noting VentureBeat reported a 92% increase in MCP servers over a three-month period.

That demand is reorganizing supply around four converging capabilities: discovery, least-privilege enforcement, behavioral oversight, and unified control planes. Microsoft’s Satya Nadella said, “Starting with observability… something called Agent 365. I need to have an inventory,” then tied inspectability and policy-governed execution to Entra, Defender, and Purview; meanwhile PR Newswire reported “NSS Labs, AWS, Microsoft, F5 Set New AI Security Standards With Two Foundational White Papers,” OWASP’s Scott Clinton used RSAC to outline evolving approaches to securing agentic systems—at RSAC Conference 2026, as Co-Chair and co-founder of the OWASP GenAI Security Project, he “shares insights from the project's latest research, including new landscape guides and evolving approaches to” securing these environments—and RockCyber Musings warned to treat Berkeley and CCIA submissions as “where auditors will focus within 18 to 36 months.” The identity layer is moving with it: Merritt Maxim discusses “how AI is affecting Identity and Access Management,” with the explicit expectation that “Expect to hear this topic a lot throughout 2026, especially as the industry tries to figure out what's” needed for agent-era access control.

Sources
Resilient CyberPR Newswire - Consumer TechnologyRockCyber MusingsApplication Security Weekly (Video)Reid Hoffman

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.