AI cybersecurity for businesses: defensive tools without the hype

What changed on the attack side

IBM's figures show AI working for attackers in two ways. First, AI makes attacks faster and cheaper to launch. Suja Viswesan, VP of IBM Security Software, summed it up: “AI is making attacks faster and cheaper, while breaches keep getting more expensive.” Second, AI systems themselves have become targets. More than 20 percent of organisations in the study reported a breach involving an AI model or application, and the most common causes were not exotic model attacks but weaknesses around them: compromised APIs, applications or plug-ins (27 percent) and cloud misconfigurations affecting AI workloads (27 percent). Cybersecurity Dive, reporting on the same study, noted that the share of incidents involving shadow AI, tools staff use without approval, more than doubled to 43 percent.

Smaller firms feel it too. In VikingCloud's 2026 SMB Threat Landscape Report, a survey of small and mid-sized business owners and cyber leaders in North America, 46 percent said they had faced AI-generated phishing in the past 12 months and 29 percent deepfake schemes. Kevin Pierce, VikingCloud's chief operating officer, warned that “Cybersecurity has crossed the line from an IT challenge to a business survival issue.”

Sophos says its X-Ops researchers tracked an attacker this year using AI agents inside a customer's network to build and test nearly 80 malware modules, alongside criminals selling brokered API keys for frontier AI models and trading jailbreak techniques on underground forums.

The October 2026 wave of AI defence tools

The new products fall into two groups. The first uses AI to run security operations. Elastic's AlertZero, announced on 8 October in technical preview, gives Elastic Security customers a team of specialised agents for triage, threat hunting, forensics and detection tuning. Notably, its Detection Watch proposes rule changes but, in Elastic's words, “will not change a rule without approval.” SentinelOne's Singularity ASOC, unveiled on 21 October at its OneCon conference, combines agentic investigation, automated response and an AI SIEM, with the company saying analysts set the limits on what the system can do.

The second group secures the AI a business already uses. Sophos AI Defense, which became generally available in October as a paid add-on for Sophos EDR, XDR and MDR customers, finds the AI tools staff are using, including unapproved ones, and applies policy to them. Kyle Falkenhagen, Sophos's senior vice president of product management, said most security teams “can't answer three basic questions: which AI tools are running, what data is going into them, and whether that activity lines up with policy.” SentinelOne's Singularity Agentic Detection and Response (ADR), launched in early access the same day as ASOC, inventories AI agents and MCP servers, blocks prompt injection and credential exfiltration, and can stop a single agent mid-action. Its chief product officer, Chris Corde, made a point every business using agents should hear: “an equal if not more prominent risk comes from unintended behavior of sanctioned agents.”

Where the hype starts

Analysts are more cautious than the launch announcements. Writing for Cybersecurity Insiders on 8 October, Rob Smith, CEO of Lionfish Tech Advisors, pointed out that vendor decks keep quoting a Gartner planning assumption that 70 percent of large security operations centres will pilot AI agents by 2028, while Gartner's own Hype Cycle for Security Operations, 2026 places AI SOC agents at the Peak of Inflated Expectations, with 1 to 5 percent market penetration and five to ten years to mainstream adoption. As Smith puts it, “A pilot is not an adoption.” He also notes that Gartner advises buyers to check what their existing SIEM and XDR vendors already have on the roadmap before buying something new.

Practitioners who have tested AI in real security teams say the same. At RSAC 2026, as reported by Dark Reading, Shilpi Mittal described using a language model as a “read-only triage assistant” in a Fortune 500 food manufacturer's SOC, with human approval gates, tool allow lists and full audit logging. Her team saw mean time to response improve by 22 percent. Ankit Gupta, who leads security at a financial company, ran a two-week test on a non-production system where AI had full control, and the results were bad: “SOC reality is messy — alerts arrive with incomplete fields, inconsistent identifiers, and ambiguous signals,” he said, adding that the “AI incorrectly removed users from the system.” His conclusion was that AI can assist, but final action calls stay with humans.

What actually helps a small or mid-sized business

The most useful number in IBM's report may be this one: organisations that used AI and automation in security operations cut breach costs by almost $2 million on average. That does not mean every business needs its own AI SOC. For most small companies, the AI already built into their email security, endpoint protection or a managed detection and response (MDR) service is the realistic way to get that benefit.

The bigger gap for most businesses is basic control over their own AI use. Among IBM's respondents that suffered an AI-related breach, 92 percent lacked proper AI access controls such as role-based access and multi-factor authentication. That is not a problem a new product solves on its own. It needs an inventory of the AI tools in use, clear rules on what data may go into them, and the same access controls every other business system should have.

A practical, hype-free checklist

• List every AI tool, plug-in and agent your team uses, including free ones staff signed up for themselves.
• Decide what data may never be pasted into public AI tools, and write it down in plain language.
• Turn on multi-factor authentication and role-based access for AI tools, admin panels and API keys.
• Give AI agents the smallest permissions they need, and require human approval for actions that delete, pay or send.
• Train staff on AI-generated phishing and deepfake voice or video requests, and agree a call-back rule for payment changes.
• Ask your current security vendors which AI features you already pay for before buying a new platform.
• For any AI security product, ask for measured results, how it fails, and what it can do without approval.
• Keep logs of what your AI tools and agents do, so you can investigate when something goes wrong.

Where CoreCodify fits

Many AI security incidents start in the surrounding software, not the model: an exposed API, a plug-in with too many permissions, a chatbot that can reach data it should not. Building AI features safely is as much about solid web and app engineering as it is about the AI itself.

At CoreCodify we design and build websites, mobile apps, UI/UX and AI solutions with access control, logging and human approval steps planned in from the start. If you want an AI feature or website built with security in mind, hire me for custom work at usamify.com.