Is AI the Root Cause of Today’s Cyberattacks?

Authors
Each time a major breach makes the news, the same question follows: was AI behind it? It is a fair question. AI tools write convincing phishing emails in any language, clone voices from short audio samples and turn a published vulnerability into working attack code far faster than a person could.
The data, though, points somewhere more useful. Attackers are using AI mostly to move faster through openings that defenders already know about, and many of those openings could have been closed before anyone exploited them. This article looks at the five gaps that matter most, what the research says about each, and the fixes that work.
What the data says about AI in attacks
Verizon’s 2026 Data Breach Investigations Report (DBIR), which covers data from October 2024 to November 2025, included a study of how threat actors use AI, built with enforcement data from Anthropic on 793 threat actors. The median actor used AI across 15 MITRE ATT&CK techniques, and only about 2.5% of that activity involved rare techniques.
Academic work points the same way. In a 2024 University of Illinois study, a GPT-4 agent given only the CVE description exploited 87% of a benchmark of 15 real-world one-day vulnerabilities, and without that description its success rate fell to 7%. The sample was small and the models have moved on since, so the numbers are indicative, but the pattern holds: AI is far better at exploiting a published flaw than at finding a new one.
Gap 1: Known flaws that stay unpatched
Exploitation of vulnerabilities is now the most common way into a breach. The 2026 DBIR puts it at 31% of breaches, up from 20% a year earlier. At the same time, the median time to fully patch grew from 32 days to 43, and only 26% of the flaws on the U.S. CISA catalog of known exploited vulnerabilities were fully remediated, down from 38%.
Verizon’s researchers say AI has helped shrink the defenders’ window from months to hours. The gap sits in that mismatch between attacker speed and patch speed, and these steps help close it:
- Prioritize by exploitation risk as well as severity. CISA’s Known Exploited Vulnerabilities catalog lists flaws attackers are already using, and CISA urges every organization to fix those first.
- Plan maintenance in advance. NIST SP 800-40 Rev. 4 recommends maintenance groups with schedules set for each risk scenario, so an urgent patch does not wait on negotiation.
- Give patching an owner and a deadline. Risk-based SLAs, named owners, pre-agreed downtime windows and a formal process for accepting risk all target the slow tail of the 43 days.
- Use AI to rank the backlog. AI can sort vulnerabilities by real-world exploitability and asset context, as covered later in this article.
Gap 2: People who can be convinced
The human element was involved in 62% of breaches in the 2026 DBIR. Email phishing remains the favorite social engineering route, and in Verizon’s phishing simulations, scams by voice or text on mobile drew a 40% higher click rate than email.
The Arup case in Hong Kong shows how far AI can take this. In early 2024 an employee received a message that appeared to come from the company’s UK-based CFO, then joined a video call where the other participants were AI-generated deepfakes. A series of transfers followed, totaling about US$25 million, and the fraud came to light only when the employee checked with headquarters. The missing control was a check through a separate channel:
- Verify high-risk requests out of band. Payments, access changes and new bank details requested by email, voice or video get confirmed through a separate, known contact before anyone acts.
- Train staff regularly. Joint guidance from CISA, the NSA, the FBI and MS-ISAC recommends regular training on social engineering and phishing.
- Use single sign-on with phishing-resistant MFA. The same guidance notes SSO reduces the chance of users handing over credentials, especially alongside MFA.
- Block known-bad infrastructure at the email gateway, using denylists of malicious domains, URLs and IP addresses.
- Write down what to do when someone thinks they have been tricked, as part of a documented incident response plan.
Gap 3: Stolen credentials that open more doors
Credential abuse was the first step in 13% of breaches in the 2026 DBIR, down from 22%, yet it appeared somewhere in 39% of breaches overall. Attackers chain a stolen identity with exploitation and privilege escalation, so weak access control makes every other gap worse.
- Move to phishing-resistant MFA. CISA calls FIDO/WebAuthn authentication the gold standard and the only widely available phishing-resistant option. SMS and voice codes are a last resort.
- Roll out in phases. Start with email, file servers and remote access, and with administrator and other privileged accounts.
- Use number matching in the meantime. For push-based MFA, CISA recommends number matching as an interim defense against MFA fatigue attacks.
- Lock accounts on unusual activity and after repeated malicious login attempts.
- Apply least privilege, so one stolen login reaches only what its owner genuinely needs.
Gap 4: AI tools nobody approved
IBM’s Cost of a Data Breach Report 2025 found that one in five organizations reported a breach linked to shadow AI, the AI tools staff use without approval, and that high levels of shadow AI added around US$670,000 to the average breach cost. Among organizations that reported a breach of AI models or applications, 97% lacked proper AI access controls, although IBM notes those breaches are still a small share of its sample.
Governance is lagging. IBM found that 63% of breached organizations either had no AI governance policy or were still developing one, and the 2026 DBIR found that 67% of users reach AI services through non-corporate accounts. Governance has to come before more tools:
- Take inventory. List the AI tools, plug-ins and built-in AI features people actually use. Nobody can govern what they cannot see.
- Write a clear AI usage policy that says which data types, such as customer data, source code, contracts and personal information, may go into which tools.
- Offer approved alternatives. People turn to shadow AI because it helps them work, so a sanctioned tool with proper contracts and data protection removes much of the reason to go around the rules.
- Apply access controls and data protection, including single sign-on, role-based access and data loss prevention rules for AI tools.
- Monitor, train and assign an owner. Watch for unsanctioned AI traffic, explain to staff what is risky and why, and name one person accountable for AI governance.
Gap 5: Weaknesses inside AI applications
The OWASP Top 10 for LLM Applications 2025 ranks prompt injection first. Because it exploits how language models process instructions, it cannot simply be patched away. The list also expanded its entry on excessive agency, where an AI agent holds more tools or permissions than its task needs. Both risks grow as assistants connect to email, databases and APIs.
- Treat model output as untrusted input. Validate inputs, filter outputs and enforce critical controls outside the model, whatever the prompt says.
- Limit what agents can do. Give each agent only the permissions its task requires, and require human approval for consequential actions.
- Keep secrets out of prompts. Handle access to external tools in code, with the application’s own scoped credentials.
- Let downstream systems decide. The systems an AI connects to should check what is allowed rather than trust the model’s judgment.
- Log and audit every prompt and action, with tamper-proof records that trace back to a user.
Where AI helps the defense
The same speed works for defenders, as long as high-risk actions still go through people and tested playbooks. AI takes on the repetitive, high-volume work so security staff can spend their time on decisions that need judgment, and four areas stand out:
- Faster detection with less noise. Machine learning scores traffic, logins and endpoint behavior against a baseline, surfaces anomalies that fixed rules miss and summarizes alerts for analysts.
- Smarter patch prioritization. AI ranks a large backlog by real-world exploitability and asset context, which helps answer the question of what to patch first. It is a ranking aid, and people still make the call.
- Phishing and fraud detection. AI analyzes suspicious messages and spots campaign patterns faster than manual review, while analysts verify high-impact and unfamiliar threats.
- Lower cost per breach. IBM’s 2025 report found that organizations using security AI and automation extensively saved an average of US$1.9 million per breach and shortened the breach lifecycle by 80 days.
What to do now, wherever you operate
AI raises the stakes on the basics, and the organizations that handle this period well will be the ones that close known gaps quickly, know who owns each risk and keep control of the AI they use. Five priorities apply across markets:
- Fix known weaknesses faster. Exploitation is now the top entry point and patching is falling behind, which makes it the first place to invest, especially for companies with complex supplier chains and legacy systems.
- Make cybersecurity a leadership responsibility. NIST’s Cybersecurity Framework 2.0 added a Govern function in 2024 that places cyber risk with leadership. Named owners, leadership dashboards and formal risk acceptance put it into practice.
- Govern AI before adding more of it. ISO/IEC 42001, the international standard for AI management systems, gives a structure to build on, and national guidance is moving the same way, including Japan’s AI Guidelines for Business, which reached Version 1.2 in March 2026.
- Prepare for stricter reporting. The EU’s NIS2 Directive asks covered entities for an early warning within 24 hours of a significant incident, Australia has required ransomware payment reports within 72 hours since 30 May 2025, and Japan enacted its Active Cyber Defense law in May 2025. Confirm what applies to your business with legal counsel and official sources.
- Use AI for defense with people in charge. Put AI to work on alert triage, patch prioritization and phishing detection, keep human approval for high-risk actions, and plan analyst capacity on the work that actually exists.
So, is AI the root cause of today’s cyberattacks? It is often in the room, speeding things up. The openings it walks through, though, are mostly the ones security teams have been talking about for years, which is good news, because those are the ones within reach to close.
References
Shea, S. (2026, May 22). Verizon 2026 DBIR: 6 key takeaways for CISOs. TechTarget. Retrieved November 9, 2026, from https://www.techtarget.com/cybersecurity/news/366643420/Verizon-2026-DBIR-6-key-takeaways-for-CISOs
Muntyan, A. (2026, August 3). Verizon DBIR 2026: 10 stats that should change your security strategy. Passwork. Retrieved November 9, 2026, from https://passwork.pro/blog/verizon-dbir-2026-stats/
Fang, R., Bindu, R., Gupta, A., & Kang, D. (2024, April 17). LLM agents can autonomously exploit one-day vulnerabilities (arXiv:2404.08144). arXiv. Retrieved November 9, 2026, from https://arxiv.org/abs/2404.08144
Arghire, I. (2026, May 19). Verizon DBIR 2026: Vulnerability exploitation overtakes credential theft as top breach vector. SecurityWeek. Retrieved November 9, 2026, from https://www.securityweek.com/verizon-dbir-2026-vulnerability-exploitation-overtakes-credential-theft-as-top-breach-vector/
Positive Technologies (2026, May 26). Verizon DBIR 2026: Exploitation of vulnerabilities becomes the leading initial access vector. PT dbugs. Retrieved November 9, 2026, from https://dbugs.ptsecurity.com/news/verizon-dbir-2026-exploitation-of-vulnerabilities-becomes-the-leading-initial-access-vector-20260526
Cybersecurity and Infrastructure Security Agency (2021, November 3). BOD 22-01: Reducing the significant risk of known exploited vulnerabilities. Cybersecurity and Infrastructure Security Agency. Retrieved November 9, 2026, from https://www.cisa.gov/news-events/directives/bod-22-01-reducing-significant-risk-known-exploited-vulnerabilities
Cybersecurity and Infrastructure Security Agency (n.d.). Known exploited vulnerabilities catalog. Cybersecurity and Infrastructure Security Agency. Retrieved November 9, 2026, from https://www.cisa.gov/known-exploited-vulnerabilities-catalog
Souppaya, M., & Scarfone, K. (2022, April). Guide to enterprise patch management planning: Preventive maintenance for technology (NIST Special Publication 800-40 Rev. 4). National Institute of Standards and Technology. Retrieved November 9, 2026, from https://doi.org/10.6028/NIST.SP.800-40r4
Noto, G. (2024, May 17). Scammers siphon $25M from engineering firm Arup via AI deepfake ‘CFO’. CFO Dive. Retrieved November 9, 2026, from https://www.cfodive.com/news/scammers-siphon-25m-from-engineering-firm-arup-via-ai-deepfake-cfo/716501/
CISA, NSA, FBI, & MS-ISAC (2023, October). #StopRansomware guide. Cybersecurity and Infrastructure Security Agency. Retrieved November 9, 2026, from https://www.cisa.gov/resources-tools/resources/stopransomware-guide
Cybersecurity and Infrastructure Security Agency (2022, October). Implementing phishing-resistant MFA [Fact sheet]. Cybersecurity and Infrastructure Security Agency. Retrieved November 9, 2026, from https://www.cisa.gov/sites/default/files/publications/fact-sheet-implementing-phishing-resistant-mfa-508c.pdf
Cybersecurity and Infrastructure Security Agency (2022, October). Implementing number matching in MFA applications [Fact sheet]. Cybersecurity and Infrastructure Security Agency. Retrieved November 9, 2026, from https://www.cisa.gov/sites/default/files/publications/fact-sheet-implement-number-matching-in-mfa-applications-508c.pdf
IBM (2025, July 30). IBM report: 13% of organizations reported breaches of AI models or applications, 97% of which reported lacking proper AI access controls [Press release]. IBM Newsroom. Retrieved November 9, 2026, from https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls
OWASP Foundation (2024, November 18). OWASP Top 10 for LLM applications 2025. OWASP Gen AI Security Project. Retrieved November 9, 2026, from https://genai.owasp.org/llm-top-10/
National Institute of Standards and Technology (2024, February 26). The NIST Cybersecurity Framework (CSF) 2.0 (NIST CSWP 29). National Institute of Standards and Technology. Retrieved November 9, 2026, from https://doi.org/10.6028/NIST.CSWP.29
International Organization for Standardization (2023, December). ISO/IEC 42001:2023 Information technology - Artificial intelligence - Management system. ISO. Retrieved November 9, 2026, from https://www.iso.org/standard/81230.html
Ministry of Economy, Trade and Industry (2026, April). Guidance on the interpretation and application of civil liability in the utilization and application of AI (Ver. 1.0, provisional translation). METI. Retrieved November 9, 2026, from https://www.meti.go.jp/english/press/2026/pdf/0409_002a.pdf
European Parliament & Council of the European Union (2022, December 14). Directive (EU) 2022/2555 on measures for a high common level of cybersecurity across the Union (NIS2 Directive). EUR-Lex. Retrieved November 9, 2026, from https://eur-lex.europa.eu/eli/dir/2022/2555/oj
Rychkova, M., Adler, N., & Kallenbach, P. (2025, June 16). Pay and tell: Mandatory ransomware payment reporting obligations in force. MinterEllison. Retrieved November 9, 2026, from https://www.minterellison.com/articles/mandatory-ransomware-payment-reporting-obligations-in-force
Paganini, P. (2025, May 19). Japan passed a law allowing preemptive offensive cyber actions. Security Affairs. Retrieved November 9, 2026, from https://securityaffairs.com/178056/laws-and-regulations/japan-passed-a-law-allowing-preemptive-offensive-cyber-actions.html