Generative AI security is the combination of governance, technical controls, and testing that ensures AI use in a company, from employee assistants to agents connected to internal systems, does not become a path to data leaks, fraud, or manipulated decisions. The goal is to make use intentional, traceable, and secure.
In 2023, the Vultus blog addressed the early risks of language models in the article 2023 analysis of language model risks. Since then, AI has begun reading documents, querying internal repositories, and taking actions, and the risk has grown alongside it.
Adoption arrived before the policy
AI adoption followed the path of other technologies that deliver quick gains: people started using it before their companies defined how. Pasting client contracts into a public assistant for summaries, uploading payroll spreadsheets to generate charts, and reviewing proprietary code with unvetted tools are common situations, with no malicious intent involved.
The numbers show the scale. According to Microsoft and LinkedIn's 2024 Work Trend Index, 75% of knowledge workers already use AI at work, and 78% of those users bring their own tools, outside what the company has purchased. The Global Cybersecurity Outlook 2026, from the World Economic Forum, shows that 87% of leaders saw AI-related vulnerabilities increase over the past year.
Banning use often pushes it onto personal devices, beyond security's visibility. Governance with an approved alternative tends to work better.
Which risks are specific to AI?
Some risks are familiar: sensitive data in the wrong place, an unassessed vendor, uncontrolled access. Others arise from how models work and are cataloged in the OWASP Top 10 for LLM Applications.
In prompt injection, text hidden in an email, document, or page contains instructions the assistant follows as though they came from the user. In agents that send emails or query databases, those instructions become actions. Systems that search internal documents may disclose information to unauthorized people if searches do not respect the original permissions. Other threats include data poisoning and techniques to bypass model safeguards.
Attackers have also started using AI. In November 2025, Anthropic reported disrupting an espionage campaign in which AI agents performed 80% to 90% of the tactical work against about 30 organizations.
The workstreams of an AI security program
The Vultus methodology combines recognized references across four areas:
- Governance: a management system following the standard of ISO/IEC 42001, with a usage policy, an AI asset inventory, and risk and vendor assessments, organized by NIST AI RMF.
- Technical controls: protection of data, models, applications, agents, and infrastructure, supported by SAIF, Google's Secure AI Framework.
- Adversarial validation: AI red teaming against models and applications in production.
- Continuous operation: retesting after every significant change and residual risk indicators for leadership.
A practical starting point includes an inventory of tools in use, a brief policy on which data may go into which tools, an official alternative, testing before production, and a record of agents' actions. Brazil's General Data Protection Law (LGPD) applies fully to personal data used in prompts, knowledge bases, or training.
AI security with adversarial validation
Vultus supports companies in governing, protecting, and testing generative AI systems. Read about AI red teaming and talk to Vultus about AI use in your company.
