Google's Gemini artificial‑intelligence system penetrated the digital defences of three separate firms during a controlled security test, a company spokesperson confirmed to the BBC.

The AI model was allowed to browse the web, locate publicly available login pages and, using pattern‑recognition techniques, generate plausible credentials that granted it access to internal portals.

"Gemini demonstrated the ability to autonomously discover and exploit weak authentication," a Google representative said.

All three targets were mid‑size enterprises operating in the technology, logistics and financial services sectors. The companies were selected because they had previously consented to participate in a red‑team exercise coordinated by an independent security consultancy.

During the test, Gemini identified login forms, harvested email addresses from corporate websites and combined them with common password structures to guess passwords. In each case, the AI succeeded in entering a user account that possessed limited, but nevertheless sensitive, privileges.

How the AI achieved the breach

Google's engineers designed Gemini to operate with a "goal‑oriented" prompt, instructing it to gain access to a specified resource. The model then used its large‑scale language capabilities to craft credential guesses, iterating rapidly and adapting to server responses.

Security analysts note that the technique mirrors traditional credential‑stuffing attacks, but the AI's ability to generate context‑aware guesses in real time reduces the need for massive pre‑compiled password lists.

Industry reaction and broader implications

Cyber‑security firms warned that the demonstration signals a shift from human‑run phishing campaigns to automated AI‑driven assaults. "If an AI can learn to crack weak passwords on the fly, the threat surface expands dramatically," said a senior analyst at Mandiant.

Google, which introduced Gemini as a competitor to OpenAI's ChatGPT, said the exercise was intended to expose vulnerabilities before they could be weaponised by malicious actors.

Context and next steps

The test follows a spate of high‑profile AI‑related security incidents, including a recent incident where a language model generated phishing emails that bypassed corporate filters. Governments worldwide are beginning to draft regulations on AI safety, and the United Kingdom's Centre for Data Ethics has called for mandatory AI security audits.

Google has pledged to share the findings with the broader security community and to incorporate stronger credential‑checking safeguards into future Gemini releases. The three companies involved have begun reviewing their authentication policies, with two already planning to enforce multi‑factor authentication for all staff.

As AI capabilities accelerate, experts say organisations must treat AI‑generated attacks as a realistic threat, not a theoretical one.