Australian AI security company Kapālins has appointed Jai A Singla as Chief Technology Officer and a director. He will lead Kapālins’ technology as AI moves from answering questions to acting on its own.
Jai’s view is that the three oldest promises in security (confidentiality, integrity and availability) now have to hold for AI itself, because frontier models are becoming autonomous faster than most security programmes were built for. The UK AI Security Institute found that by February 2026 the length of cyber tasks AI could complete on its own was doubling about every 4.7 months. The choice keeps widening too: Hugging Face now lists more than 3.1 million public models, and the number keeps growing.
The three promises, applied to AI that acts
Confidentiality: What a model or agent can read, remember and pass on, and to whom.
Integrity: Whether each model and agent still acts as it was approved to, release after release.
Availability: The AI services a business now depends on stay up, within budget and under control.
Independent research shows why integrity can no longer be assumed. In one routine evaluation, run with internet access, the UK AI Security Institute recorded agents taking 19 actions outside their brief. METR found that at least 16% of frontier agents’ successes on tasks of eight hours or more were not legitimate. The OWASP Top 10 for Agentic Applications now lists memory and context poisoning, where a planted instruction shapes what an agent does long after it was written.