Skip to content

AI Cyber Attack Statistics 2026: AI-Powered Phishing, Attacks on AI Systems and Defenses

AI cyber attack statistics for 2026 from IBM, Verizon, the FBI, Microsoft, OpenAI, Anthropic and Google: AI phishing, attacks on AI systems and AI defenses.

Omar KhandakerBy , Senior Security Consultant20 min read

AI & LLM Penetration TestingFrom $4,500Fixed scope, free retest, no hourly billing.

Get my fixed price

AI shows up on both sides of an attack now. Criminals and state groups use models to write phishing, build malware and find vulnerabilities, and the AI systems companies deploy have become targets in their own right. This page collects the measured numbers on both: breach data from IBM and Verizon, FBI complaint data, Microsoft’s telemetry, the threat reports OpenAI, Anthropic and Google publish on misuse of their own models, and academic studies of AI phishing and prompt injection. Each figure links to the source that produced it.

How to cite: link to this page or to the primary source beside each figure. Survey figures carry their sample size.

Key numbers at a glance

  • 1 in 4 malicious breaches was AI-enabled in IBM’s 2026 study, a 56% increase, and those breaches cost $6 million on average against a global average of $4.99 million. (IBM Cost of a Data Breach 2026, 602 organizations)
  • 22,364 complaints to the FBI in 2025 referenced AI, with $893 million in reported losses. (FBI IC3 2025 Internet Crime Report)
  • 54% of people clicked fully AI-automated spear phishing emails, the same rate as emails written by human experts, against 12% for a generic phishing control. (Heiding et al. 2024, 101 participants)
  • More than 20% of organizations reported a breach targeting their AI models or applications. (IBM 2026)
  • 832 malicious accounts used Claude across all 14 MITRE ATT&CK tactics and 482 techniques between March 2025 and March 2026. (Anthropic LLM ATT&CK Navigator 2026)
  • 45% of employees now use AI regularly on corporate devices, up from 15% a year earlier, and 67% of users reach AI services through non-corporate accounts on corporate devices. (Verizon 2026 DBIR)
  • $1.93 million lower breach costs for organizations that use security AI and automation extensively, compared with those using none. (IBM Cost of a Data Breach Report 2026)

1. How often attackers use AI

  • 1 in 4 malicious breaches was AI-enabled, up 56% from the previous year. IBM says the rise was led by AI deepfake impersonation and AI-enabled malware. (IBM Cost of a Data Breach 2026, 602 organizations, breaches from March 2025 to February 2026)
  • AI-enabled breaches averaged $6 million, roughly $1 million above the global average of $4.99 million. (IBM 2026)
  • 62% of the AI-driven attacks in IBM’s study targeted critical infrastructure sectors. Financial services breaches averaged $6.3 million and energy breaches $5.2 million. (IBM 2026)
  • A year earlier, 16% of breaches involved attackers using AI, most often for phishing or deepfake impersonation. (IBM Cost of a Data Breach 2025, 600 organizations)
  • AI-enabled adversaries increased their operations 89% year over year in 2025. CrowdStrike tracks more than 280 named adversaries. (CrowdStrike 2026 Global Threat Report)
  • The median threat actor in Verizon’s data researched or used AI assistance in 15 documented attack techniques, and some used it in as many as 40 or 50. (Verizon 2026 Data Breach Investigations Report, dataset October 2024 to November 2025)
  • The FBI received 22,364 complaints with an AI reference in 2025, with adjusted losses above $893 million. (FBI IC3 2025)

Note on terms: IBM’s “AI-enabled breach” means the attacker used AI. That is different from a breach of an AI system, which sections 6 to 8 cover.

2. AI-generated phishing and social engineering

  • 54% of participants clicked fully AI-automated spear phishing emails, 54% clicked emails written by human experts, 56% clicked AI emails with a human in the loop, and 12% clicked the generic control. The authors estimate AI can make phishing up to 50 times more profitable against large audiences. (Heiding, Lermen, Kao, Schneier and Vishwanath, arXiv 2024, 101 participants)
  • 82.6% of phishing emails KnowBe4 analyzed between September 15, 2024 and February 14, 2025 showed some use of AI, and polymorphic tactics appeared in 76.4% of campaigns. (KnowBe4 Phishing Threat Trends Report, Vol. 5, 2025)
  • In March 2025, Hoxhunt’s AI phishing agent was 24% more effective than its human red team, after being 31% less effective in 2023. This is the vendor’s own test (70,000 AI-generated simulations), and the page also gives the result as 23.3% and 23.8%. (Hoxhunt 2025)
  • Phishing was the initial access vector in 23% of the intrusions Microsoft’s incident responders handled from July 2025 to June 2026, up from 7% a year earlier. (Microsoft Digital Defense Report 2026)
  • 62% of organizations experienced a deepfake attack involving social engineering or exploiting automated processes in the previous 12 months. (Gartner survey, September 2025, 302 cybersecurity leaders)
  • Businesses reported more than $30 million in 2025 losses to business email compromise involving AI. Investment fraud with an AI nexus passed $632 million, romance scams more than $19 million, voice-cloning “distress” scams more than $5 million and AI-involved employment scams almost $13 million. (FBI IC3 2025)
  • Since April 2025, criminals have impersonated senior US officials with text messages and AI-generated voice messages. (FBI IC3 Public Service Announcement I-051525-PSA, 2025)
  • AI-generated TikTok videos in a 2025 ClickFix campaign, some with about 500,000 views, used synthetic narration to talk viewers into installing the Vidar or StealC infostealers. (Microsoft Digital Defense Report 2026)

The click-rate study matters more than the volume figures: AI-written spear phishing performed like a skilled human attacker, at a fraction of the cost. Training programs built around spotting typos and generic greetings are testing for the wrong thing. Our phishing attack statistics page has the wider click-rate and business email compromise data.

3. State actors and criminal groups using commercial AI models

OpenAI, Anthropic and Google publish reports on accounts they have banned. Each sees only its own platform, but these are the most direct evidence of how attackers use AI.

  • 832 accounts associated with malicious cyber activity used Claude for all 14 ATT&CK tactics and 482 unique techniques between March 2025 and March 2026, in 13,873 observed actions. (Anthropic, Mapping AI-enabled cyber threats, June 2026)
  • 69% of those actors (574 of 832) used AI to develop capabilities, and 560 used it for malware development. Only 6.5% used it for lateral movement and 22.5% for privilege escalation and impact. (Anthropic, June 2026)
  • The share of actors rated medium risk or higher for AI enablement rose from 33% to 56% in under a year. (Anthropic, June 2026)
  • A group Anthropic assessed with high confidence to be Chinese state-sponsored used Claude Code against roughly 30 organizations in September 2025, with AI performing 80% to 90% of the campaign and humans stepping in at about 4 to 6 decision points. It succeeded in a small number of cases. (Anthropic, November 2025)
  • One criminal used Claude to run a data extortion campaign against at least 17 organizations, including healthcare and emergency services, with ransom demands that sometimes exceeded $500,000. Another sold ransomware developed with Claude for $400 to $1,200. (Anthropic, Detecting and countering misuse of AI, August 2025)
  • In cases Anthropic disrupted between December 2025 and August 2026, AI-assisted breaches were completed in two to three hours, with individual operators handling dozens of victims in parallel. One criminal group mass-downloaded 1.8 million Android apps in a credential-harvesting pipeline and exfiltrated more than a terabyte of data. (Anthropic, Detecting and countering misuse of AI, September 2026)
  • A suspected Russian state-nexus actor targeted more than 20 organizations in its operational planning, and a separate actor targeted roughly 50 organizations and produced more than a dozen possible zero-day findings in a single month. (Anthropic, September 2026)
  • OpenAI has disrupted and reported more than 40 networks that violated its usage policies since February 2024. In the cyber cases, it found no evidence that its models gave attackers new tactics or novel offensive capabilities. (OpenAI, Disrupting malicious uses of AI, October 2025)
  • Government-backed groups from more than 20 countries used Gemini: more than 20 China-backed groups, more than 10 Iran-backed, 9 North Korea-backed and 3 Russia-backed. APT42 accounted for more than 30% of Iranian APT use. (Google Threat Intelligence Group, January 2025)
  • One model extraction attempt against Gemini used more than 100,000 prompts. (Google Threat Intelligence Group, February 2026)

Read together, the provider reports describe a shift. Through 2025, Google (January) and OpenAI (October) described AI as a productivity aid that gave attackers no novel capability. From November 2025, Anthropic documented AI running most of an intrusion and AI-assisted breaches finishing in hours. Anthropic’s own data also shows where AI is used least: lateral movement and privilege escalation inside a network.

See what a real pentest finds. Read our sample report, then get a fixed quote in writing.

4. AI-assisted malware

  • In 2025, Google identified the first malware families that query a large language model while running, listing five families including PROMPTFLUX and PROMPTSTEAL. APT28’s use of PROMPTSTEAL against Ukraine was the first malware querying an LLM that Google saw in live operations. (Google Threat Intelligence Group, November 2025)
  • AI-assisted malware development mostly reproduces known techniques: the median AI-assisted function had 55 existing malware examples doing the same thing, and fewer than 2.5% of observations involved less-common techniques. (Verizon 2026 DBIR)
  • The s1ngularity supply chain attack (August 2025) hijacked locally installed AI command-line tools such as Claude Code, Gemini CLI and Amazon Q CLI to hunt for secrets, leaking about 2,000 secrets and 20,000 files across 225 victims. (Microsoft Digital Defense Report 2026)

Malware that asks a model what to do next generates different behavior on each run, so indicators from one incident carry over poorly to the next. Microsoft’s report says the useful signal shifts to behavior: unexpected traffic to inference endpoints and AI command-line tools run with permissive flags.

5. AI vulnerability discovery, for attackers and defenders

  • Google identified, for the first time, a criminal using a zero-day exploit it believes was developed with AI, a two-factor authentication bypass planned for mass exploitation. (Google Threat Intelligence Group, May 2026)
  • The median time from a vulnerability’s discovery in the wild to its weaponization is now well below 24 hours, and a record of about 72,000 CVEs is expected to be tracked in 2026. (Microsoft Digital Defense Report 2026)
  • In a lab test, Anthropic’s Mythos and OpenAI’s GPT-5.5 achieved full domain compromise of an emulated enterprise through a 32-step attack chain with no human help. Open-weight models trail closed models in orchestrating attacks by about seven months. (Microsoft Digital Defense Report 2026)
  • Exploitation of vulnerabilities became the most common initial access vector for breaches, at 31%, while credential abuse fell to 13%. (Verizon 2026 DBIR)
  • Claude Opus 4.6 found and validated more than 500 high-severity vulnerabilities in open-source software. (Anthropic, February 2026)
  • Claude Mythos Preview found thousands of high-severity vulnerabilities, including in every major operating system and web browser, among them a 27-year-old bug in OpenBSD. Anthropic committed up to $100 million in usage credits for defenders. (Anthropic Project Glasswing, April 2026)
  • In DARPA’s AI Cyber Challenge final, autonomous systems found 86% of synthetic vulnerabilities and patched 68% of those found, discovered 18 real vulnerabilities, analyzed more than 54 million lines of code and spent about $152 per task. (DARPA AIxCC results, August 2025)
  • XBOW’s autonomous tester reached first place on HackerOne’s US leaderboard after submitting nearly 1,060 vulnerabilities; over 90 days, program owners rated 54 critical and 242 high severity. (XBOW, June 2025)
  • Google’s Big Sleep agent found a SQLite vulnerability (CVE-2025-6965), the first time an AI agent was used to directly foil efforts to exploit a vulnerability. (Google, July 2025)

The same capability serves both sides. If weaponization takes under a day, a quarterly patch cycle leaves months of exposure for anything internet-facing. Anthropic’s February 2026 analysis questions whether industry-standard 90-day disclosure windows can hold against the volume of LLM-discovered bugs. (Anthropic, February 2026)

6. Attacks on AI systems: prompt injection and AI agents

  • More than 20% of organizations reported a breach targeting AI models or applications. The most common causes were in the systems around the model: compromised APIs, applications or plug-ins (27%) and cloud misconfigurations affecting AI workloads (27%). (IBM 2026, 602 organizations)
  • In 2025, 13% of organizations reported breaches of AI models or applications and 8% did not know. Of those breached, 97% had no AI access controls; 60% of the incidents compromised data and 31% disrupted operations. (IBM 2025, 600 organizations)
  • 32% of organizations experienced an attack on an AI application that used the application prompt in the previous 12 months. (Gartner survey, September 2025, 302 cybersecurity leaders)
  • Attackers injected malicious prompts into legitimate generative AI tools at more than 90 organizations in 2025. (CrowdStrike 2026 Global Threat Report)
  • In Microsoft’s telemetry from Azure AI workloads over a 90-day window in 2026, malicious link injection through model and tool outputs made up 52% of attack activity, jailbreaks 16%, tool and agent abuse 12%, identity and credential abuse 10%, and prompt injection and instruction subversion 8%. (Microsoft Digital Defense Report 2026)
  • By goal, 45% of observed malicious prompts sought policy-violating use, 17% sensitive data exfiltration, 14% system prompt leakage, 12% goal hijacking of downstream tools and 11% leakage of system internals. (Microsoft Digital Defense Report 2026)
  • 88% of enterprises are experimenting with AI agents and 82% of leaders plan broader rollouts within 12 to 18 months. (Microsoft Digital Defense Report 2026)
  • In a public red-teaming competition against 22 frontier AI agents in 44 scenarios, participants submitted 1.8 million prompt injection attacks and more than 60,000 caused policy violations. Nearly all agents violated policy for most behaviors within 10 to 100 queries. (Zou et al., arXiv 2025)
  • Against an adaptive attacker with 100 attempts per environment, Claude Opus 4.5 in browser use was successfully prompt-injected 1.4% of the time with improved safeguards, down from 8.0% for the model alone. Anthropic notes that a 1% rate still represents meaningful risk. (Anthropic, November 2025)
  • EchoLeak (CVE-2025-32711), an AI command injection flaw in Microsoft 365 Copilot that allowed information disclosure over a network, was rated 9.3 (critical) by Microsoft. (NVD, 2025)
  • Prompt injection is ranked LLM01, the first entry, in the OWASP Top 10 for LLM Applications 2025, followed by sensitive information disclosure and supply chain. (OWASP 2025) OWASP released a separate Top 10 for Agentic Applications on December 9, 2025. (OWASP 2025)

The IBM causes are the useful part: most breaches of AI applications came through ordinary APIs, plug-ins and cloud configuration, not exotic attacks on the model. Our explainer on how integrations expand the LLM attack surface shows how that happens, and indirect prompt injection covers the instructions hidden in documents and web pages that agents read. For agents with tools, see agentic AI security; each OWASP category is explained with test cases in our OWASP Top 10 for LLM Applications guide.

7. AI infrastructure, model supply chain and stolen AI keys

  • 2,130 AI-related CVEs were disclosed in 2025, a 34.6% increase over the previous year. (TrendAI State of AI Security Report, March 2026)
  • Nearly 230,000 internet-facing servers hosted Ollama at some point over four months, and more than 113,000 answered basic status queries. Researchers also found 2,500 exposed Chroma vector databases and more than 35,000 exposed ComfyUI instances. (TrendAI 2026)
  • 1,467 Model Context Protocol (MCP) servers were exposed to the internet, and MCP servers had 95 CVEs in 2025 after almost none before. (TrendAI 2026)
  • About 100 malicious machine learning models were found on Hugging Face. (JFrog, 2024)
  • In late March 2026, the criminal group TeamPCP claimed supply chain compromises of popular GitHub repositories including LiteLLM, an AI gateway, and monetized stolen credentials through ransomware and extortion groups. (Google Threat Intelligence Group, May 2026)
  • One extortion group attacked roughly 30 AI companies in about four days, with demands of $1.5 million to $2.5 million. (Anthropic, September 2026)
  • Nearly all AI cloud subscriptions attacked in Microsoft’s 2026 data involved generative AI services; non-generative services such as speech and vision accounted for less than 2%. (Microsoft Digital Defense Report 2026)
  • LLMjacking, the theft of cloud credentials to resell access to hosted models, could cost a victim more than $46,000 a day in model consumption if undiscovered. (Sysdig, 2024) In one 2026 case, an attacker used a stolen API key for about three weeks. (Anthropic, September 2026)

Every exposed model server, vector database and MCP server in the TrendAI counts is an inventory and access-control failure that an external test finds quickly. Our MCP server security testing guide covers what to check, and the LLMjacking and stolen AI keys write-up covers recent breaches that started with exposed keys.

8. Shadow AI and data leakage to AI tools

  • 45% of employees are regular AI users on corporate devices, up from 15%, and 67% of users access AI services with non-corporate accounts on corporate devices. (Verizon 2026 DBIR)
  • Shadow AI became the third most common non-malicious insider action in Verizon’s data loss prevention data in 2025, a fourfold increase. Source code was the data type most often submitted to external models, and 3.2% of violations involved research and technical documentation. (Verizon 2026 DBIR, 858,440 DLP events)
  • The average organization saw 223 generative AI data policy violations a month, double the year before; the top quarter saw 2,100. (Netskope Cloud and Threat Report 2026, data October 2024 to October 2025)
  • Of sensitive data sent to AI apps, 42% was source code, 32% regulated data and 16% intellectual property. 50% of organizations lack enforceable data protection policies for generative AI apps. (Netskope 2026)
  • Prompts sent to SaaS AI apps grew sixfold in a year, from 3,000 to 18,000 a month in the average organization, while the share of AI users on personal AI apps fell from 78% to 47%. (Netskope 2026)
  • 1 in 5 organizations reported a breach due to shadow AI, which added $670,000 to breach costs. 63% of breached organizations had no AI governance policy or were still writing one, and only 34% of those with a policy audit for unsanctioned AI. (IBM 2025, 600 organizations)
  • A malicious browser extension with more than 600,000 installs harvested ChatGPT and DeepSeek conversation histories, affecting almost 10,000 organizations before it was stopped. (Microsoft Digital Defense Report 2026)

Chat history is now a target because it holds what employees paste into it: source code, architecture notes, customer data and credentials. The Verizon and Netskope numbers show use rising fast. Approved tools with enterprise accounts and data loss prevention on the prompts are the controls the data supports.

9. AI in defense

  • Organizations that used security AI and automation extensively saved $1.93 million per breach compared with those using none. One in four organizations has still not adopted these tools in security operations. (IBM Cost of a Data Breach Report 2026; IBM newsroom 2026)
  • In 2025, extensive security AI saved an average of $1.9 million and shortened the breach lifecycle by 80 days. (IBM 2025)
  • More than 50% of organizations use AI agents for threat detection and containment, but only 18% apply them to vulnerability management. (IBM follow-on study, 2026, 456 organizations)
  • 85% of organizations plan to increase security spending after learning of advanced frontier AI models, up from 64% in the initial research. (IBM follow-on study, 2026, 456 organizations)
  • Security Copilot users report summarizing threats 60% to 70% faster; Microsoft’s phishing triage agent saves security teams nearly 200 hours a month; and Microsoft’s own security team uses AI agents to investigate 75% of incoming incidents. (Microsoft Digital Defense Report 2026)
  • ChatGPT is used to identify scams up to three times more often than it is used to run them, by OpenAI’s estimate. (OpenAI, October 2025)

IBM’s 18% figure is the gap worth noting. Defenders apply AI to detection, while attackers apply it to finding and exploiting vulnerabilities. Finding your exploitable weaknesses first, in the AI stack and around it, is the job of an AI and LLM penetration test: it covers prompt injection and agent tool abuse, and the APIs, plug-ins and cloud configuration that IBM’s data says cause most AI breaches.

What to expect in 2027

These are the sources’ own forecasts and trends drawn from the data above, labelled as such.

  • Fully automated intrusions (Microsoft forecast). Microsoft expects fully automated agentic breaching systems to start continuously assessing internet-exposed enterprise assets, and expects the volume of complex attacks to grow as AI orchestration lets attackers scale. It cites an industry projection of about 1.3 billion AI agents in production by 2028. (Microsoft Digital Defense Report 2026)
  • Shorter patch windows (trend). Weaponization in under 24 hours (Microsoft), exploitation as the top breach vector at 31% (Verizon) and AI models finding hundreds to thousands of high-severity bugs (Anthropic) all point the same way, and Anthropic says industry-standard 90-day disclosure windows may not hold. (Microsoft 2026; Verizon 2026 DBIR; Anthropic 2026)
  • More attacks on AI systems themselves (trend). Breaches of AI models and applications rose from 13% of organizations in IBM’s 2025 study to more than 20% in 2026, and AI-related CVEs rose 34.6% in 2025. (IBM 2025; IBM 2026; TrendAI 2026)
  • Changing security approaches (survey). 67% of cybersecurity leaders say emerging generative AI risks demand significant changes to existing security approaches. (Gartner, September 2025, 302 leaders)

Part of our cybersecurity statistics hub. See also phishing attack statistics and data breach statistics.

Sources

Frequently asked questions

How many cyber attacks use AI? In IBM’s 2026 study of 602 breached organizations, one in four malicious breaches was AI-enabled, up 56% in a year, and those breaches cost $6 million on average. CrowdStrike measured an 89% rise in operations by AI-enabled adversaries in 2025. No source counts every AI-assisted attack; these figures cover studied breaches and tracked adversaries.

Is AI-generated phishing more effective than human phishing? It matches skilled humans. In a 101-person study, fully AI-automated spear phishing got a 54% click rate, the same as emails written by human experts and well above the 12% of a generic phishing email. The authors estimate AI can make phishing up to 50 times more profitable at scale.

How are state hackers using ChatGPT, Claude and Gemini? Mostly for research, coding, phishing content and malware development. Google saw government-backed groups from more than 20 countries using Gemini, and OpenAI has disrupted more than 40 networks since February 2024. Anthropic’s analysis of 832 malicious accounts found 69% using AI to build capabilities, and in one 2025 espionage campaign AI did 80% to 90% of the work against about 30 targets.

How often are AI systems themselves breached? More than 20% of organizations reported a breach targeting AI models or applications in IBM’s 2026 study, up from 13% in 2025. The most common causes were compromised APIs, applications or plug-ins (27%) and cloud misconfiguration (27%). In 2025, 97% of organizations with an AI breach had no AI access controls.

What is shadow AI and how common is it? Shadow AI is employees using AI tools the organization has not approved. Verizon found 45% of employees now use AI regularly on corporate devices, and 67% of users reach AI services through non-corporate accounts. IBM’s 2025 study found one in five organizations had a breach due to shadow AI, adding $670,000 to the cost.

Does AI help defenders? Yes, measurably. Organizations using security AI and automation extensively saved $1.93 million per breach in IBM’s 2026 study, and in 2025 they contained breaches 80 days faster. Only 18% of organizations use AI agents for vulnerability management, though.

Omar Khandaker

Written by

Omar Khandaker, Senior Security Consultant

11+ years in the field. Focus: PCI DSS, ISO 27001, ISO 9001 & CMA. Part of the team that scopes, tests, and reports for our clients. Full profile →

Find out what an attacker sees.

Tell us what to test and see your fixed price.

Prefer the full scoping questionnaire? 
Start the conversation

Get a quick proposal

We are online