Google: threat actors are using AI across the attack lifecycle

Google Threat Intelligence Group (GTIG) reports that threat actors are using the company's Gemini model across several stages of an attack, from reconnaissance through to post-compromise activity. The report identifies three trends in late 2025: model extraction, where actors systematically query a model to replicate aspects of its logic and behaviour, AI-augmented operations, and early attempts to build AI into tooling itself. State-backed actors are using large language models for target profiling, translation, open-source collection, phishing content and coding. GTIG notes that frontier models have not yet been directly targeted by advanced persistent threat groups, but that the experimentation argues for continued monitoring and strong API protections.

What this means for your organisation

The short-term effect is not new forms of attack, but lower cost and higher tempo in the familiar ones. Fluent Norwegian phishing, aimed at named individuals in your organisation, no longer costs the attacker much effort. That undermines every control resting on employees spotting clumsy language.

Berigo recommends

• Shift security training away from linguistic tells and towards context and action: unexpected urgency, changed account numbers, requests outside the normal channel.
• Require out-of-band confirmation for payments and changes to supplier details, however convincing the request appears.
• Protect your own AI APIs with rate limiting, authentication and logging, so systematic querying becomes visible.
• Bring AI-assisted social engineering into incident response exercises, not just the awareness campaign.

Source

• cloud.google.com