A new
OWASP report on GenAI security incidents shows cyberattacks rapidly evolving from theoretical risks to real-world exploitation, with attackers increasingly leveraging AI to scale operations and target critical systems.
Among the most significant cases is
a breach of Mexican government agencies, where attackers used AI tools such as
Anthropic’s Claude and ChatGPT to automate reconnaissance and exploit development.
The campaign exposed roughly 150 GB of sensitive tax and voter data, demonstrating how AI can compress the time and effort required to execute large-scale intrusions. Researchers said the attack expanded across multiple agencies, highlighting the growing risk to public-sector systems as AI accelerates attack workflows.
Risks in cloud-based AI infrastructure
In one case, researchers uncovered a
“Double Agent” scenario in
Google Cloud’s Vertex AI platform, where an overprivileged agent could abuse default permissions to access sensitive data, extract credentials, and pivot into broader cloud resources.
The findings underscore concerns around identity and privilege management in agentic systems, particularly as organizations rely more heavily on managed AI services with complex trust boundaries.
Supply chain vulnerabilities are another growing concern
A breach involving
AI data vendor Mercor, linked to
compromised versions of the LiteLLM tool, raised fears that proprietary training data workflows and contractor information across major AI labs could have been exposed. The incident prompted Meta to pause work with the vendor and highlighted how third-party dependencies in AI ecosystems can introduce cascading risks across multiple organizations.
Across these incidents, OWASP researchers emphasize a broader shift in the threat landscape. Rather than focusing solely on model-level vulnerabilities, attackers are exploiting weaknesses in identity, orchestration layers, and interconnected supply chains. Many of these risks stem from misconfigurations, excessive permissions, and design flaws rather than traditional software bugs.
The report concluded that securing AI systems will require a more holistic approach, combining system-level controls, stronger identity governance, and tighter oversight of third-party dependencies as AI becomes a central component of both enterprise and government operations.
Read about more trends from the OWASP GenAI roundup of incidents from the first quarter of 2026.
