ESET’s latest readiness analysis argues for simpler controls, automated triage and human expertise around the decisions that matter.
Introduction
AI is increasing both business capability and the number of systems attackers can target. ESET’s WeLiveSecurity analysis says SMB adoption is moving faster than governance and argues that security needs to become simpler to operate, not merely larger. The article cites ESET research indicating that 73 percent of SMBs are integrating AI and 40 percent of businesses lack a proper AI policy. It also describes risks in AI skill repositories and prompt injection. These figures come from a cybersecurity vendor and its research, so buyers should treat them as a signal rather than neutral proof of a particular product. The practical conclusion is credible: small teams need fewer unmanaged tools, clearer ownership and automation supervised by people who understand business context.
What happened
ESET reports that it analysed 900,000 skills across several AI-agent repositories between March and May 2026, identifying more than 25,000 as suspicious and over 3,000 as malicious. The article says some could exfiltrate data, execute malware, abuse permissions or alter agent behaviour. It also notes that AI systems with access to sensitive data and business applications enlarge the attack surface, especially when permissions are excessive. According to the cited readiness index, nearly half of global SMBs suffered a cybersecurity incident in the previous year, while many lack regularly reviewed response plans. The publication advocates automation for repetitive analysis and prioritization, combined with human experts for ambiguous investigations and consequential decisions.
Why it matters for growing businesses
An agent skill resembles a plugin: it extends capability, but it may also receive access to files, credentials, web services or commands. Installing one from an unfamiliar repository can create the equivalent of giving a stranger a key and hoping the description is accurate. Prompt injection adds another path because malicious instructions can hide in emails, documents or web pages that an agent reads. SMEs may be more exposed than larger firms because one administrator often manages endpoints, cloud accounts, ERP and backups. The answer is not to ban every extension or purchase disconnected security products. It is to know what is installed, limit what each component can reach and make detection and containment routine.
Practical actions this week
Inventory AI applications, browser extensions, agent skills and automation connectors on company accounts and devices. Record source, version, owner, permissions and business purpose. Remove items with no owner or clear need. Allow installation only from reviewed sources and pin versions where possible. Use separate service identities instead of employee administrator accounts, and default integrations to read-only. Protect email, cloud storage, Odoo, banking and code repositories with strong multifactor authentication. Centralize endpoint, identity and cloud alerts so somebody is accountable for reviewing them. Patch internet-facing systems quickly and test offline or immutable backups. For every agent, define what data it may read, which actions require approval and how its credential can be revoked without disabling the entire business.
A practical review cycle
Run a monthly access review for high-risk agents and a quarterly review for the rest. Compare configured permissions with actual tool calls; unused write access should be removed. Sample logs for unusual destinations, bulk reads, failed authentication and activity outside working patterns. Re-test prompt-injection cases after model or skill updates. Confirm that alerts reach a named person and that the person can isolate a device, revoke a token and pause an automation. Conduct one restore test and one incident tabletop each quarter. Keep the evidence in a simple register that management can understand. Security maturity is not the number of alerts collected. It is the speed and confidence with which the team can identify, contain and recover from a real problem.
How VDS Consultoria can help
VDS Consultoria can map the effective access of AI tools across endpoints, cloud platforms, automation services and Odoo. We identify abandoned extensions, shared credentials and integrations with permissions broader than their jobs. We then build a proportionate control set: managed identities, approved repositories, logging, review cadence, backup tests and response procedures. Where an SME lacks round-the-clock capacity, we help select monitoring and support services around clearly defined responsibilities rather than another stack of unattended consoles. The goal is resilient automation that reduces repetitive work while keeping sensitive systems observable and recoverable. Secure AI adoption is ordinary security discipline applied to faster, more connected software.
Leadership checklist
Ask for a current list of agents, skills and connectors, including their owners and credentials. Confirm who reviews alerts and who has authority to stop an automated workflow. Check whether important systems use unique accounts, strong authentication and tested recovery. Require evidence that permissions match the task, not the maximum offered during installation. Review supplier security, update practices and data handling before connecting production information. Measure time to detect, contain and restore, alongside operational uptime. If the business cannot answer what an agent can reach, the first project is discovery, not expansion. Capability without inventory is simply surprise with a monthly subscription.
Source
Cyberthreats are moving faster than SMBs: Readiness must accelerate, ESET WeLiveSecurity, 2026-09-16.