When AI Becomes the Attack Vector | Security Medic Consulting

You’ve heard that AI is changing everything about cybersecurity, and you’re not sure whether to be scared or prepared. Both instincts are reasonable. This article explains what’s actually happening, why it matters for your organization specifically, and five concrete things you can do about it this week — without a security team.

§01

The attack has been automated

You probably heard “we’re too small to be a target” and believed it. Artificial intelligence just made that idea obsolete — not by making attackers smarter, but by making targeting free.

There is a particular kind of organizational belief that costs small businesses and public agencies everything. It sounds like this: “We’re too small to be a serious target.” Cybersecurity professionals have been arguing against that idea for twenty years. AI just made it obsolete in a new and specific way — and most people running small organizations haven’t fully absorbed what that means yet.

Here is what actually changed. Attackers used to choose targets manually. They picked organizations worth their time — large enough to pay a ransom, valuable enough to compromise. That took labor, research, and skill. Today, a single attacker using AI tools can simultaneously target every small business in a zip code, generating personalized attack messages for each one at near-zero marginal cost. The economics of targeting have collapsed. You are not too small to be targeted. You are exactly the size attackers want: real data, real money, and defenses that haven’t kept pace with the threat.

That’s the threat in plain terms. The next section explains how these attacks actually work mechanically — not to frighten you, but because understanding the mechanism is what turns a vague worry into something you can act on.

What the numbers say IBM’s 2025 Cost of a Data Breach Report put the global average breach cost at $4.44 million — and in the U.S., a record $10.22 million. One in five breaches now involves shadow AI: unsanctioned AI tools your employees are already using without IT’s knowledge, adding $670,000 to the average cost when they become breach vectors. For a dental practice, a nonprofit, or a county agency, a breach at this scale is frequently not recoverable.[1]

IBM’s 2025 report also found that 16% of all data breaches now involve attackers using AI — most often through AI-generated phishing (37% of AI-driven attacks) and deepfake impersonation (35%). Generative AI has compressed the time to craft a convincing, personalized phishing email from 16 hours to five minutes.[1] If you’ve ever received an email that seemed a little too specific to be generic spam, you’ve already encountered this technology.

§02

Four ways AI is being used against you right now

These aren’t future scenarios. Each of these attack types is active today and has been documented against small and mid-sized organizations.

AI is being used across the full lifecycle of an attack — from finding you, to deceiving your people, to exploiting your systems, to hiding inside them once they’re in. Here is what each phase actually looks like under the hood.

Phase 1

AI-Powered Reconnaissance

Before a human attacker types a single command, automated AI tools have already built a profile of your organization. They scan your publicly visible software stack for known vulnerabilities, probe your cloud storage for misconfigured buckets, scrape LinkedIn for employee names and roles, and map vendor relationships from your website. By the time a human attacker reviews the output, your organization is already fully characterized. The attack starts before you know it’s started.

Phase 2

In 2024, a finance employee at an international firm wired $25 million after a video call with what appeared to be her CFO and several colleagues. The entire call was AI-generated[4] — voice, face, and context synthesized from publicly available recordings and images. The employee had no way to tell. This attack type targets any organization where one person can authorize a financial transaction — which describes almost every small business.

Phase 3

Automated Vulnerability Exploitation

AI tools cross-reference publicly disclosed software vulnerabilities with the technology stack identified during reconnaissance, then automatically generate working exploit code. The gap between a vulnerability being announced and attackers actively using it — already measured in days rather than weeks — is compressing further. Unpatched systems are no longer a theoretical risk. They are an active one, and AI is accelerating the timeline.

Phase 4

Adaptive Malware

Once inside a system, AI-assisted malware can analyze the security tools it encounters and rewrite its own code to avoid triggering them. Traditional antivirus relies on recognizing known threat signatures — if the malware changes its signature, traditional tools miss it entirely. This is why attackers can remain inside networks for weeks or months before detection. What gets in through this mechanism tends to stay in much longer than most organizations expect.

16% Of all breaches involved attackers using AI (IBM 2025)

$10.2M Average U.S. breach cost — record high (IBM 2025)

97% Of AI-related breaches involved systems without proper access controls

63% Of organizations have no AI governance policy in place

That 97% figure is worth pausing on. Nearly every AI-related breach IBM studied involved systems that lacked proper access controls.[1] Not advanced defenses. Basic controls: who can access which systems, and how do you verify it’s really them. This is not a resource problem. It’s one every organization can address.[1]

§03

The governance gap is the real vulnerability

The threat intelligence reports focus on attacker technology. Two decades of security practice makes something else clear: governance — not technology — is what decides whether an organization survives a breach.

Most security articles center on attacker capability. After two decades of security practice, what stands out is something different: the technical sophistication of the attack is rarely the deciding factor in whether an organization survives. The organizations that recover from AI-enabled attacks are almost never the ones with the most advanced security tools. They are the ones that knew what systems they had, knew who was responsible for protecting them, and had practiced what to do when something went wrong.

IBM’s data makes the governance problem concrete. Among organizations that experienced an AI-related breach, 97% lacked proper AI access controls. Shadow AI — the ChatGPT, Copilot, Gemini, and other AI tools your employees are already using without telling IT — accounted for 20% of all breaches. Shadow AI incidents took a full week longer to detect, added $200,000 to average breach costs, and compromised customer PII in 65% of cases compared to 53% globally. The tools your people adopted without a conversation are the ones creating your exposure right now.[1]

“The breach is not coming from the AI tool you evaluated. It is coming from the three your team adopted without telling anyone.”

The constructive reality embedded in all of that: this is fixable. You don’t need a security operations center to address shadow AI. You need a conversation with your team, a policy that gives them clarity, and an inventory of what tools are actually in use. That’s governance. It’s not a technology purchase. It’s a set of decisions and conversations that any person leading an organization can make, regardless of technical background.

The NIST CSF 2.0 connection The questions in the next section map directly to the six functions of the NIST Cybersecurity Framework 2.0[3]: Govern, Identify, Protect, Detect, Respond, Recover. The “Govern” function — new in CSF 2.0 — explicitly places cybersecurity risk at the leadership level, recognizing it as a business risk, not just a technical one.[3] You don’t need to read the full framework. You just need to be able to answer the questions below honestly.

The Verizon 2024 Data Breach Investigations Report[2] frames it from a different angle: most breaches are not sophisticated nation-state operations. They are opportunistic. Attackers probe for unlocked doors, and they move on if they don’t find one quickly. Governance is how you lock the doors. It’s also how you know which doors exist.[2]

§04

What you need to know as the person in charge

This used to be framed as guidance for “the CEO,” “the CISO,” and “the CTO.” That framing assumes you have a team with those titles. You probably don’t. Here is the same guidance organized by the concerns that actually keep you up at night.

Whether you’re running a dental practice, managing a school district, directing a nonprofit, or overseeing operations for a small manufacturer, you wear many hats. Security leadership is one of them now, whether or not you signed up for it. These are the questions you should be able to answer — or know exactly who to call when you can’t.

Concern: Access & Identity

Who can get into what — and how do I know it’s really them?

Does every external-facing system require multi-factor authentication (MFA)? When did you last review which employees have access to which systems — and remove access for people who no longer need it? What happens to accounts when someone leaves the organization? These three questions address the single most effective control against AI-enabled attacks. Most breaches exploit accounts with excessive access and insufficient verification — often accounts that belonged to former employees and were never disabled.

Concern: People & Awareness

Would my team recognize an AI-generated deception?

AI-generated phishing emails are grammatically perfect and contextually specific — they reference your actual vendors, your staff by name, your current projects. Voice cloning can produce a call that sounds exactly like your CFO or your IT director. Video deepfakes can make a video meeting look entirely convincing. Has your team seen real examples of these attacks? Do they have a clear procedure to verify unusual requests — especially financial ones — through a second, independent channel? A call back on a known number, not a reply to the original message, is the standard protocol.

Concern: Visibility

What AI tools are my people already using?

Shadow AI is already in your organization. Employees adopt tools that help them work faster, and they don’t always tell anyone — not because they’re being deceptive, but because it doesn’t occur to them that it matters. Take an inventory: what AI tools are being used, by whom, and what data do those tools send to external servers when they’re used? This isn’t about restricting your team. It’s about knowing your exposure so you can manage it deliberately. IBM found that 63% of organizations have no AI governance policy. Yours can start as a single page.[1]

Concern: Detection

If something went wrong 30 days ago, would I know today?

This is one of the most clarifying questions in security. The average time between breach and detection is measured in weeks or months — not hours. Attackers stay quiet intentionally, moving slowly to avoid triggering alerts while they map your systems, escalate their access, and position for maximum impact. Do you have any mechanism to notice unusual login locations, unexpected file access patterns, or large data transfers happening outside business hours? You don’t need enterprise tools to start. Basic logging and alerting built into Microsoft 365 or Google Workspace can surface early signals if it’s turned on and someone is looking.

Concern: Response & Recovery

If we got hit tomorrow, what would we do in the first 24 hours?

This question has a name in security: an incident response plan. It doesn’t need to be long. It needs to answer: who gets called first, which systems get isolated, who communicates with customers or patients or parents, and who contacts legal counsel and cyber insurance. Organizations with a documented plan — even a simple one — recover faster, spend less on recovery, and make fewer costly decisions under pressure. Organizations without one improvise while under stress, and the outcomes are consistently worse and more expensive.

Concern: Data & Vendors

What happens if our data gets stolen, not just encrypted?

Modern ransomware attacks typically follow a two-step sequence: exfiltrate the data first, then encrypt it. This means the threat is no longer just “pay us to restore access to your files.” It’s “pay us, or we publish your patient records, client contracts, student information, or financial data.” Do you know what data you’re holding, where it lives, and which third-party vendors have access to it? Data minimization — retaining only what you actually need, for as long as you need it — directly reduces what an attacker can threaten to expose.

§05

Five things you can do this week

None of these require a security team. All of them meaningfully reduce your exposure. Start wherever feels most accessible and build from there.

You now have a clear picture of what’s happening and why it matters. The primary vulnerability in most AI-enabled breaches is not a technology gap — it’s a governance gap: not knowing what systems you have, who has access to them, and what your organization will do if something goes wrong. These five actions address that gap directly. Pick one today and finish it this week.

01 Audit your access controls. List every system your organization uses — email, file storage, practice management software, financial tools, payroll — and verify that multi-factor authentication is enabled on every external-facing one. Then check who has access that no longer needs it. For anything that holds patient, client, student, or financial data, this is the single highest-return security action you can take this week. The 97% figure from IBM’s research starts here.
02 Ask your team what AI tools they use. Keep it short and non-punitive — a 15-minute team meeting or a three-question survey. You’re not trying to catch anyone doing something wrong. You’re trying to understand what data is leaving your organization through tools nobody vetted. Shadow AI is already present in your organization. The only question is whether you know what it is.
03 Brief your team on deepfake social engineering. Show them the Hong Kong video call example — the one where a finance employee wired $25 million to what appeared to be her CFO on a video call. Make the mechanism real and concrete. Then establish one simple rule: any unusual financial request — wire transfers, gift card purchases, changes to banking information — requires verbal verification on a known phone number, completely separate from the original request. This single protocol has stopped many expensive incidents.
04 Test your backups. Many organizations run backups and never verify they work. Right now, restore a single file or folder from backup. If that process is unclear, takes more than 30 minutes to figure out, or fails, fixing it is urgent. A working, tested backup is the difference between a disruptive week and a catastrophic one when ransomware hits. Knowing the backup works before you need it is the definition of this step.
05 Write a one-page incident response plan. Who is the first call if you suspect a breach? Who communicates with customers, patients, or parents? Who handles media inquiries? Who contacts legal counsel? Who notifies cyber insurance? Writing those names and phone numbers down — before anything happens — is consistently underrated as a security measure. It doesn’t need to be sophisticated or lengthy. It needs to exist and to be findable when someone needs it under pressure.

The bigger picture These five actions map directly to the NIST Cybersecurity Framework 2.0[3] functions: Identify (knowing your systems and data), Protect (MFA and awareness training), Detect (understanding what normal looks like), Respond (the incident plan), and Recover (verified backups). The “Govern” function — the newest addition to CSF 2.0 — is exactly what you’re doing by working through these steps: making security decisions at the leadership level, treating cybersecurity risk as business risk, and not delegating it entirely to whoever manages the technology.

The organizations that recover from AI-enabled attacks are not the ones with the most sophisticated technology. They are the ones that knew what systems they had, who was responsible for protecting them, and what to do in the first hours when something went wrong. They are the ones where someone in leadership treated security as a business function — not an IT expense, not a compliance checkbox, not someone else’s problem.[3]

AI changes the threat. It does not fundamentally change what a sound security posture looks like. What AI does is remove the protection that small organizations once had simply by being too small for manual targeting. That protection is gone. The time it would have taken an attacker to profile your organization, craft a convincing approach, and find a way in — that time has collapsed to near zero. The good news is that practical readiness is within reach for any organization willing to work through the right questions. You have them now. The next step is answering them.

Security Medic Consulting

Not sure where your organization stands? Let’s find out together.

Security Medic Consulting works with small businesses, medical practices, nonprofits, and SLED organizations across the Hudson Valley. We translate security risk into plain language and build governance frameworks that work for the size and budget you actually have — not the enterprise you’re not.

Start the Conversation

References

IBM Security & Ponemon Institute. (2025). Cost of a Data Breach Report 2025: The AI Oversight Gap. IBM Corporation. https://www.ibm.com/reports/data-breach
Verizon. (2024). 2024 Data Breach Investigations Report (DBIR). Verizon Business. https://www.verizon.com/business/resources/reports/dbir/
National Institute of Standards and Technology. (2024). The NIST Cybersecurity Framework 2.0. U.S. Department of Commerce. https://doi.org/10.6028/NIST.CSWP.29
Hong Kong Police Force. (2024, February). Deepfake video call fraud — HKD 200 million loss. South China Morning Post. https://www.scmp.com/news/hong-kong/law-and-crime/article/3250851/
Cavoukian, A. (2009). Privacy by Design: The 7 Foundational Principles. Information and Privacy Commissioner of Ontario.
CISA. (2023). Shifting the Balance of Cybersecurity Risk: Principles and Approaches for Security-by-Design. CISA. https://www.cisa.gov/resources-tools/resources/secure-by-design

Security Medic Consulting, LLC · Hudson Valley, New York ← All Articles