AI Is Accelerating: It’s Time To Address Governance, Visibility And Security
James C. Foster is the CEO of eSentire.
gettyMany organizations now have an agentic workforce. AI agents are everywhere, reading emails, writing code and connecting to systems on our behalf, analyzing financial reports, creating dashboards and doing the work for your human workforce.
Unfortunately, in most businesses, security teams can’t tell you which ones are running, what they can reach, where they send data or what they’re costing us.
Attackers have also built their own agentic workforce. In Ocean’s Eleven, pulling off one heist took a crew of 11 specialists: someone to recon the target, someone to crack the vault and someone to drive. Today, a single operator can hand most of those roles to AI agents that work around the clock and never ask for a coffee break.
The numbers bear this out. Microsoft’s 2026 Digital Defense Report found that the median time from a vulnerability’s discovery in the wild to its weaponization has fallen to well below 24 hours. In late 2025, Anthropic documented an espionage campaign in which AI executed an estimated 80% to 90% of tactical operations independently. The company characterized it as “the first documented case of a cyberattack largely executed without human intervention at scale.”
The cases are piling up, too. In July 2026, a report was published on JadePuffer, which it assessed as the first documented case of agentic ransomware. The agent broke into an exposed server, harvested credentials, pivoted to a separate production database and encrypted more than 1,300 configuration items. When one step failed, it deployed a working fix in 31 seconds. In September 2026, the Dutch Institute for Vulnerability Disclosure (DIVD), an organization that hunts bugs for a living, was breached by an agentic AI attacker who chained two zero-days in the institute’s help desk software and achieved root access in seconds.
A crew like that changes the economics of cybercrime. Persistence is automated, reconnaissance runs in parallel and failed tactics are retried instantly. In my view, the limiting resource is increasingly compute, and lower-cost models are making that resource cheaper for attackers every quarter.
Defenders need the same speed, with one key difference: accountability. Purpose-built agents should carry the high-volume, repetitive work while people keep authority over consequential decisions.
That balance only holds when organizations can set the rules for their agents, understand what those agents are doing and protect them from being turned against the business. AI usage and adoption at machine speed rests on three tenets: governance, visibility and security.
Governance starts with decision rights. Before an agent touches production, define what it may do by itself: Execute a specific fix, isolate an endpoint or revoke a user session. Then define how it acts within these rights.
Tie that authority to the criticality of what the agent touches. An agent quarantining an entry-level employee’s laptop is probably a non-event; an agent taking a payment server offline is a disruption you created yourself. Adaptive guardrails scale an agent’s freedom to the stakes, with a human in the loop exactly where the stakes demand one.
Every good crew still has someone who calls the big moves. In the campaign Anthropic disrupted, humans stepped in at only a handful of decision points, such as approving active exploitation and deciding how much data to take. If attackers reserve their biggest calls for people, defenders should be at least as deliberate.
But this must be ongoing. Due to the speed of agentic advancement, guardrails reviewed once a quarter are out of date before the meeting ends.
No heist crew moves before it cases the target. Treat agents like the workforce they are, each with an identity, an owner and a budget. Few leaders can list every agent running in their environment, the privileges each one holds, the data it can reach, the actions it is authorized to take and what it costs to run.
JadePuffer showed the stakes. Its way in was an AI development tool: an internet-facing Langflow server with a flaw patched more than a year earlier. The agent mined it for stored credentials and hunted for keys to commercial AI services, which let attackers run their agents on someone else’s bill. Every AI tool an organization stands up is both an asset and an attack surface.
Visibility also means seeing how findings connect. A vulnerability scanner may flag one weakness, an identity tool another and the security operations center a third, without anyone noticing that together they form a viable attack path.
The key metric here is time-to-engage: the time from the first meaningful signal to active investigation or containment.
Heist movies always need an insider, and every agent you deploy could become one, too.
Instructions hidden in an email, document or web page can turn an agent into a super-agent with broad access. You need to secure and manage agents like the privileged users they are, with least-privilege credentials, monitored access to tools and data and defenses against hidden instructions.
Then you need to secure the stack that builds and runs them. AI frameworks, model connectors and agent platforms are reaching production faster than most patching and review processes can follow. Hold them to production standards: Patch them promptly, keep them off the open internet unless they need to be there and review every new agent connection before it goes live.
Layered defense still works. DIVD said network segmentation and its incident responders kept the attacker from going deeper, and the attack was loud and messy enough to spot. Agentic attackers are fast. They lose to defenders able to watch and contain at machine speed.
AI will keep accelerating, and attackers will keep assembling their own crews. The organizations that come out ahead will treat governance, visibility and security as the foundation for moving fast: deciding what agents may do on their own, knowing every one that is running and protecting them like the privileged workforce they are.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
