Article written by Sigurd Mathiesen, Board Member at Finterai.

Purpose

The project “Phishport – Your Own Security Expert” aimed at developing a scalable
prototype for realtime identification and analysis of phishing attacks. Its goal was to
strengthen cybersecurity for both project partners and external organizations, while
supporting compliance with international standards and EU cybersecurity regulations.
The project was carried out in collaboration between Finterai AS and Sparebank 1 - Sør
Norge , with Finterai’s CEO Alexander Skage serving as project manager. It was conducted
with support from the Research Council of Norway.

Execution

The project focused on developing AI-based phishing detection, building the Phishport
prototype, and performing testing and validation. The dataset consisted of opensource
material and synthetic phishing messages, and the collaboration with Sparebank 1 - Sør
Norge provided valuable insights into practical usage.
Deliveries progressed as planned. However, the team chose not to implement a sandbox for
multimodal AI models, partly because affordable websitechecking services already exist and
because better methods are needed to defend against LLM jailbreaking.

Results

The project’s main objectives were largely achieved. The prototype for realtime phishing
identification demonstrated strong performance, with a high detection rate and few false
positives. An external sandbox solution was validated for safe opening of links and
attachments.

Although ceratin preparations for commercialization remain, the project has established a
strong foundation for continued development. It has contributed to improved cybersecurity for
external organizations through courses, guidance materials, and enhancements to
cybersecurity products. The banking and finance sector has especially benefited from this
initiative.

The project also supported compliance with international standards and EU cybersecurity
regulations. Its' scope covered:

- information security
- application security
- laws and regulations
- standards
- risk assessment
- threat modeling
- secure data storage and transmission
- cloud and hybrid environment security
- AI systems
- supply chains
- cybercrime
- phishing
- privacy
- digital trust

Additionally, the project strengthened security awareness and competence among
employees.

Key technical results included:

- A method for detecting when language models behave correctly versus incorrectly
during phishing detection — highly relevant for defense against advanced threat
actors and LLM jailbreaking.
- Very high accuracy in phishing-email detection, with minimal false positives and false
negatives.
- Tests on adversarial phishing and prompt injection demonstrated that modern
language models are increasingly robust.
- Efficiency and improved security can be achieved by distilling large language models
into smaller ones and through context engineering.

Conclusion

Finterai has demonstrated that generative AI can be used effectively with easy access
through technical integrations in applications to protect against phishing, thus providing
significant societal value for both individuals, banks and other relevant businesses.