ESET shields its AI agents: cybersecurity starts billing for the autonomy the industry itself is selling

By Zendoric · October 2, 2026.

ESET has added to PROTECT, its B2B security platform, a module specifically aimed at organizations already using autonomous AI agents. According to the company itself, it includes two pieces: a protection layer against "malicious AI components" within the agent ecosystem, and a continuous monitoring system that tracks those agents' behavior to detect suspicious actions before they escalate into an incident. It also renames its AI Advisor assistant as LiveAI and strengthens LiveCortex, the engine that correlates and responds to threats across the platform. The company says it is allocating an additional 40 million euros to AI during 2026, a figure it attributes to Michal Jankech, its vice president of Enterprise & SMB/MSP, though the statement does not detail how that investment is distributed nor offer customer numbers or independently measured effectiveness.

The announcement should be put in its place: it is a product note, not a technical study. There are no benchmarks, no adoption data, no examples of real attacks stopped by the tool. What is significant is the context in which it appears. In the same outlet that carried this announcement, just days earlier, Proofpoint unveiled an almost identical move —unifying AI and data security under a single umbrella— and Kaspersky published recommendations that same day for closing the cybersecurity talent gap. This is no coincidence: the entire corporate security sector is racing to create, almost simultaneously, the same product category.

That category has a working name among industry analysts: "agent security" or agentic behavior governance. And it responds to a very concrete problem we have been flagging at Zendoric for months: when an AI stops merely generating text and starts acting —executing code, moving files, calling APIs, making decisions with real permissions inside a company— the attack surface changes in nature. Filtering harmful content at a model's input or output is no longer enough; you have to watch what the agent does while it operates, with which credentials, and whether that behavior deviates from what is expected. ESET acknowledges this explicitly when it justifies the launch by pointing to "new attack surfaces arising from the adoption of AI tools by employees, developers and organizations".

Our reading is that announcements of this kind say little taken one by one; taken together, they say quite a lot. They confirm that the short-term risk of agentic AI is not a hypothetical runaway superintelligence, but something far more prosaic and already present: agents with access to corporate systems that can be manipulated, misuse their permissions or simply fail silently, on a scale no human team can supervise line by line. It is the same phenomenon we pointed to when discussing how agentic AI industrializes fraud: the immediate danger is operational, not existential.

There is also an underlying paradox worth underlining. The same companies —and often the same vendors— that sell the productivity promise of autonomous agents are the ones that now have to sell, separately, the security to contain them. It is a cost that rarely appears in the initial ROI calculation of an AI deployment, and it confirms an idea we have already repeated: investing in AI without first preparing the organization's governance does not save money, it shifts it to a later bill, usually a more expensive one. In that sense, this kind of product is not a marketing whim on ESET's part; it is the necessary plumbing —unglamorous but indispensable— for agent adoption not to stall out of mistrust.

In the long run, that plumbing is precisely what makes possible the leap we advocate: an economy where autonomous agents free up human time for judgment, relationships and the work that really matters, instead of remaining a constant source of incidents and mistrust. But that horizon only arrives if the oversight layer matures at the same pace as the autonomy layer. For now, what we have is the start of a race among several vendors to put a name and a price on a real problem; whether those tools work as promised is something that can only be verified with third-party data, not with press releases.