Sovra Recon·Guide · AI & Cybersecurity

Cybersecurity and AI: How Artificial Intelligence is Changing Offense and Defense

By Solan Barlet, co-founder of Sovra Recon
Updated October 29, 2026

In short

Artificial intelligence is transforming cybersecurity on both sides. Attackers use it to craft credible phishing campaigns, accelerate vulnerability research, and automate their attacks. Defenders use it to detect anomalies, triage alerts, and, increasingly, to attack their own systems before hackers do, using offensive security agents. For a company, the implication is clear: the time between the disclosure of a vulnerability and its exploitation is shrinking, so testing must be done more frequently.

How Attackers Use AI

  • Phishing and social engineering: error-free, personalized messages translated into any language, or even imitated voices or videos (deepfakes) used in CEO fraud schemes.
  • Vulnerability research: AI helps read code, understand a published patch, and deduce how to exploit unpatched versions.
  • Automation: reconnaissance, credential testing, payload adaptation—tasks that once required human time are now performed at scale.
  • Lowering the barrier to entry: less-skilled attackers are achieving results that were previously reserved for experienced professionals.

ANSSI has dedicated several publications to these uses; its findings align with those of major software vendors: AI primarily amplifies existing techniques rather than inventing new ones.

How Security Professionals Use AI

  • Detection: identifying abnormal behavior in logs, on the network, or on endpoints (EDR, XDR, SIEM).
  • Alert triage: summarizing an incident, correlating signals, and reducing SOC team fatigue.
  • Code security: assisted code review, detection of secrets and vulnerable dependencies right from the development stage.
  • Offensive security: AI agents that conduct real penetration tests to find vulnerabilities before an attacker does.

AI-powered offensive security: attacking to protect

This is the most tangible change for small and medium-sized businesses. A manual penetration test requires expert time and often takes place only once a year—while the application itself changes every week.

Offensive AI agents can conduct this testing in a matter of hours and repeat it regularly. The goal is not to generate more alerts, but to identify proven vulnerabilities—those that an attacker could actually exploit—complete with proof-of-concept and remediation.

Securing AI Itself

Companies that integrate language models (chatbots, assistants, agents connected to their data) are opening up a new attack surface. The OWASP Top 10 for LLM Applications lists the main risks:

  • Prompt injection: Malicious text hijacks the model’s instructions, either directly or through a document it reads.
  • Sensitive information leaks: the model reveals data from its instructions, its knowledge base, or other users.
  • Excessive agent privileges: An agent has overly broad permissions (such as sending emails or modifying data) and can be manipulated to exploit them.
  • Uncontrolled Outputs: The model’s response is injected as-is into a page or a command.

These applications are tested just like any others: with a defined scope, authorization, and real exploitation attempts.

What European regulations say

The European Artificial Intelligence Regulation (AI Act, Regulation 2024/1689) requires high-risk AI systems to meet an appropriate level of accuracy, robustness, and cybersecurity, including protection against manipulation attempts such as data poisoning or inputs designed to deceive the model (Article 15). Its requirements will be phased in between 2025 and 2027.

For companies subject to NIS2, AI systems are part of the information system that must be protected: they are included in risk analysis, vendor management, and security testing.

What should be done in practice?

  • Reduce the time to patch: Apply critical patches within days, not months, prioritizing vulnerabilities that are actively being exploited.
  • Testing more frequently: moving from an annual penetration test to testing after every major update.
  • Secure authentication: Implement MFA everywhere, prioritizing email, VPN, and administrative accounts.
  • Raise awareness about AI-assisted fraud: implement a two-step verification process for all wire transfers or changes to bank account information.
  • Mapping AI usage: which tools are used, what data is entrusted to them, and what permissions agents have.

Frequently asked questions

Does AI make cyberattacks more dangerous?

This makes attacks faster, more numerous, and more credible—especially when it comes to phishing. Basic security measures (patches, MFA, backups, regular testing) remain effective but must be implemented more quickly.

What is an AI agent in cybersecurity?

A program based on a language model that pursues a goal autonomously: it chooses its tools, observes the results, and adapts its actions. In offensive security, it conducts penetration tests; in defense, it triages alerts or investigates incidents.

How can you test the security of a chatbot or an AI application?

With a dedicated penetration test that covers prompt injection, data leaks, agent privileges, and integrations—in addition to standard application testing.

Sources

Find out where an attacker would get in.

Our AI agents actually attack your scope, fully automated, and hand you the proof of every vulnerability.