Infosecurity Magazine reports that agentic artificial intelligence systems are most threatened by corrupted context, dynamic tool sourcing and supply chain risks, and authentication and authorization errors.
Most significant of the risks was context corruption, which could occur across various channels, including document repositories, chat histories, and other AI output, reported by Protect AI Chief Architect Sean Morgan at Black Hat USA 2025. AI agents are also made vulnerable by the flexibility of combining assets through the Model Context Protocol, while interactions between various tools and services introduce supply chain threats. Mounting permission transitions in AI agents should also be addressed through the use of AI-specific security tools that enable end-to-end visibility, which allow precise identity and authorization mapping, monitoring, and control, according to Morgan. Agentic AI developers have been urged by Morgan to advance security by ensuring internal and software-as-a-service agentic workload awareness, instruction context control and visibility, and proper authentication and authorization controls in protocols and systems, as well as testing their agentic AI and SaaS solutions.
Most significant of the risks was context corruption, which could occur across various channels, including document repositories, chat histories, and other AI output, reported by Protect AI Chief Architect Sean Morgan at Black Hat USA 2025. AI agents are also made vulnerable by the flexibility of combining assets through the Model Context Protocol, while interactions between various tools and services introduce supply chain threats. Mounting permission transitions in AI agents should also be addressed through the use of AI-specific security tools that enable end-to-end visibility, which allow precise identity and authorization mapping, monitoring, and control, according to Morgan. Agentic AI developers have been urged by Morgan to advance security by ensuring internal and software-as-a-service agentic workload awareness, instruction context control and visibility, and proper authentication and authorization controls in protocols and systems, as well as testing their agentic AI and SaaS solutions.
