In its State of Generative AI 2025 report, Palo Alto Networks details the rising adoption of genAI tools across enterprise, based on observations from tens of thousands of customer tenants.

The report reveals a sharp increase in genAI use – traffic surged more than 890% in 2024 alone – as businesses integrate tools for writing, coding, conversation, and more – for better or worse.

But with adoption comes mounting security risks, particularly from writing assistants and autonomous AI agents, the report warns.

The report opens with a vision of modern AI integration in the workplace: AI tools that listen in on meetings and auto-generate minutes, or AI agents managing the entire sales process from outreach to follow-up.

These scenarios are no longer hypothetical.

Companies are pursuing genAI at a pace likened to the early days of cloud computing, but with even more disruptive potential. Some organisations, the report notes, are already seeing productivity gains that yield up to 40% in incremental economic impact.

This drive toward genAI is not limited to custom development. Businesses are embracing both off-the-shelf and in-house AI tools for their rapid deployment, fixed costs, and advanced feature sets.

Today, the average organisation uses 66 GenAI applications, though this can range from a few tools to hundreds.

Adoption often accelerates around new AI model releases. Following the launch of DeepSeek-R1 in January 2025, for instance, DeepSeek traffic spiked by 1,800% over the next two months.

The most common uses for genAI fall into a few key categories:

• Writing assistants: 34%

• Conversational agents: 28.9%

• Enterprise search: 10.6%

• Developer platforms: 10.3%

Together, these four account for 83.8% of all genAI activity. Their popularity stems from their ability to address essential, high-volume tasks such as drafting emails, offering real-time query responses, retrieving relevant enterprise data, and aiding software development.

Lesser-used categories include:

• Audio generators: 5.6%

• Productivity assistants: 3.3%

• Code generators: 3%

• Meeting assistants: 2.1%

These tools, while valuable for niche tasks, haven’t reached the same level of daily use.

GenAI tools frequently serve multiple roles, the report states. For example, enterprise search and conversational agents often use the same foundational technologies – such as large language models and natural language processing – allowing for both information retrieval and real-time interaction.

Grammarly and Microsoft Copilot Dominate Usage

Among genAI apps, Grammarly leads with 39.59% of transactions, followed by Microsoft 365 Copilot at 13.52%, Microsoft Power Apps at 12.64%, and ChatGPT at 8.24%. Others in the top 10 include DeepL, Google Duet AI, Bing AI, Notion, Tabnine, and Codeium.

Grammarly’s could fairly easily attritubed to its success in establishing integration into platforms like Microsoft Word, Google Docs, and email clients.

However, its widespread use also raises risks. Palo Alto notes that writing assistants inherently require access to user content – emails, documents, and potentially confidential material – opening the door to data leakage, security vulnerabilities, and inappropriate or offensive outputs.

A separate security study found that over 70% of tested writing assistants were susceptible to jailbreaking methods that produced harmful or inappropriate responses, such as instructions for illegal activity or offensive content. The report stresses that organisations must implement strong security guardrails to prevent violations and misuse of these powerful tools.

Microsoft Copilot: Leading the AI Agent Frontier

Conversational platforms like ChatGPT, Gemini, and Microsoft Copilot are evolving into AI agents capable of carrying out autonomous, multistep tasks. These agentic systems understand instructions, execute plans, access external tools, and make decisions with minimal input.

Palo Alto’s data reveals that 49% of organisations now use Microsoft Copilot or Copilot Studio. Integrated across Microsoft 365, Copilot accesses emails, calendars, chats, documents, and more to deliver personalised results, while Copilot Studio enables companies to build bespoke AI agents.

However, these advanced AI agents bring significant operational risks.

Attackers can manipulate prompts to trigger unauthorised access or actions, and agents with elevated permissions may be exploited for malicious activities. Additionally, overreliance on AI-generated outputs without proper oversight can lead to errors or security breaches.

Security testing of both open-source and packaged AI agents uncovered vulnerabilities such as the exfiltration of cloud service account tokens, manipulation of the agents’ memory causing flawed decisions, and unauthorised access to sensitive data through agent tools.

High-Tech and Manufacturing Industries Drive AI Coding Tool Use

AI developer tools such as Microsoft Power Apps, GitHub Copilot, Hugging Face, Tabnine, and Codeium are transforming how software is built, offering real-time suggestions and automating coding tasks.

The High Technology and Manufacturing sectors lead usage, accounting for nearly 39% of all coding-related GenAI transactions. Tech firms use AI to improve software quality and speed; manufacturers benefit from AI-driven design, prototyping, demand forecasting, and quality control.

Yet the risks mirror those of other GenAI platforms:

• Sensitive data exposure: some AI tools may store proprietary code without adequate safeguards.

• Malicious code generation: insecure or exploitable code patterns can be introduced by AI.

• Legal liability: questions around IP ownership and responsibility for generated code remain unresolved.

Commenting on the report’s finding, Sam Rubin, SVP of Unit 42 at Palo Alto Networks said: “The latest OpenAI report reinforces what Unit 42 has been tracking across real-world investigations: AI is already being weaponized by threat actors.

Recommended reading

• Digital Immaturity Stalling UK AI Ambitions
• Data Skills Gap Leading to Productivity Shortfalls
• AI Skills Gap Puts Public Sector Projects in Jeopardy

“We’re seeing it leveraged in ransomware negotiations, North Korean-linked deepfakes, internal copilots used for reconnaissance, and highly tailored phishing campaigns. These aren’t future risks—they’re happening now, and we have the examples to prove it.

“Attacker use of LLMs is accelerating, and as these models become more advanced, we can expect attacks to increase in speed, scale, and sophistication. It’s no surprise that threat actors—from profit-driven cybercriminals to state-sponsored groups like those aligned with China—are embracing LLMs.

“They lower the barrier to entry and dramatically improve the believability of malicious content. In one model we tested, 51 out of 123 malicious prompts slipped past safety filters—a 41% failure rate that makes it clear today’s guardrails aren’t holding the line.

“Unit 42 has jailbroken models with minimal effort, producing everything from malware and phishing lures to even instructions for creating a molotov cocktail.

“This is the new reality of cyber-threats. AI is reshaping the attacker toolkit—and defenders must adapt just as quickly.”