Blog
AI security research & insights
MAS MindForge MAS MindForge practitioner guide: operationalising AI governance for Singapore FSI
A practitioner guide to MAS Project MindForge — FEAT Principles, 17 Considerations, seven risk dimensions, GenAI evidence, and a 90-day playbook for Singapore financial institutions.
Jailbreaking How LLMs get jailbroken, and what actually works as defense
DAN prompts, multi-turn escalation, encoding attacks, and adversarial suffixes. How LLMs get jailbroken and the defenses that actually hold up.
LLM Security 8 ways hackers are targeting your LLMs
Prompt injection, jailbreaking, model stealing, and five more LLM attack techniques every CISO needs to understand.
OWASP OWASP LLM Top 10 for 2025: What changed and why it matters
OWASP LLM Top 10 2025 adds System Prompt Leakage, Vector Weaknesses, and Unbounded Consumption. What changed and what enterprises must do now.
Multi-modal New attack vectors in multi-modal AI
Why vision, audio, and document pipelines expand the attack surface, and how to test them without treating safety as text-only.
Evaluation Benchmarking LLM guardrails
Move beyond vanity metrics: how to measure refusal quality, utility loss, and robustness the way a motivated attacker would stress your stack.
Industry Trends The State of AI Security 2025
How frameworks, incidents, and buyer expectations shifted once agents and RAG became default architecture, not science projects.
Trust layer for enterprise AI