AI Security Incidents forensics report
A year or two ago we constantly heard the same question literally from everyone. "yeah, AI can be vulnerable but are there real incidents already?" we almost ended up paying cybercriminals to perform one ... oh cmon! just joking ;)
Now especially last month news were bombarded with quite profound AI security incidents from Asana AI to the latest Amazon Q. It felt like someone in the galaxy knew that we are preparing a report and decided to add a few lines ;)
Long story short, we at Adversa AI just released the Top AI Security Incidents (2025 Edition) — a detailed, real-world report that shows how AI systems are actually failing in practice, and what security teams can do to stop it. Show it to your CISO next time if he will ask about real AI attacks.
From chatbots leaking personal info, to agents triggering unauthorized crypto transfers, to cross-tenant data exposure via MCP — the incidents are real, and the impact is serious.
↓ Some key points:
— Prompt injection caused 35.3% of all documented cases.
— Even simple prompts led to $100K+ in real losses.
— GenAI was involved in 70.6% of incidents — but Agentic AI caused the most dangerous ones.
— Many failures happened outside the model — in Models, memory, MCP and the logic between components.
What’s inside:
- Vivid visualizations to expose where AI systems are failing — by time, type, sector, and severity.
- Data Across Layers: Timelines, exploit complexity matrices, and stack-wide failure maps reveal how attacks evolve
- 17 real examples, clearly explained.
- A look at how the attacks worked.
- Steps teams can take to prevent similar issues.
References to AI Security initiatives from OWASP GenAI Security Project , Coalition for Secure AI, Cloud Security Alliance, MITRE ATLAS.
Steve Wilson, Christina Liaghati, PhD, Daniele Catteddu, Omar Santos, David LaBianca, Ken Huang, CISSP, Anton Chuvakin, Kapil R. R., Michael Coates
→ Download the full report:
https://lnkd.in/dcFASFcG
If you’re building, deploying, or protecting AI systems — this report is for you.
Let me know what you think or feel free to share with others who might find it useful.
#AIsecurity
#GenAISecurity
#AgenticAISecurity
#CISO
#AIRedTeaming
#Cybersecurity