Institute briefing
Agentic AI Risk: Cisco Updates AI Security Framework
Cisco has updated its AI security and safety framework to address the emerging risks of agentic AI systems, which can act and adapt independently. This highlights the need for businesses to move beyond traditional cybersecurity to dynamic risk management for autonomous AI.
What happened
Cisco has updated its internal framework for AI security and safety to address the emerging risks associated with "agentic AI" systems. This revision acknowledges that as AI models become more autonomous and capable of independent action, the traditional security paradigms focused on data and model integrity are insufficient. The updated framework now incorporates considerations for how these agentic systems interact with their environment, make decisions, and potentially generate new risks without direct human intervention at every step.
Why it matters
This development is significant for AI governance and business strategy as it highlights a critical evolution in AI risk. The shift from static models to dynamic, agentic systems introduces new vectors for security vulnerabilities and ethical dilemmas. Businesses deploying or developing AI must now consider not only the security of their data and algorithms but also the potential for autonomous AI agents to deviate from intended behaviour, misuse access, or generate unforeseen consequences. This necessitates a proactive approach to risk management that extends beyond conventional cybersecurity, encompassing the entire lifecycle and operational context of AI agents.
The Institute take
The industry's focus on agentic AI risk underscores a fundamental challenge in responsible AI: the increasing autonomy of these systems demands a parallel evolution in our oversight mechanisms. Many organisations are still grappling with basic AI governance, yet the frontier is already moving towards AI that can act and adapt independently.
Organisations must move beyond static risk assessments to dynamic, continuous monitoring of AI agent behaviour. The core principle here is 'controllability by design' – building in safeguards and human oversight points from the outset, rather than attempting to retroactively impose controls on highly autonomous systems. Leaders should be asking not just "Is our AI secure?" but "Can we confidently predict and control the actions of our AI agents in novel situations?" The answer to the latter is far more complex and requires a new level of AI literacy among decision-makers.
Briefing notes
Questions this story answers
01What is agentic AI?+
Agentic AI refers to AI systems that are autonomous and capable of independent action, making decisions and interacting with their environment without constant direct human intervention.
02Why did Cisco update its AI security framework?+
Cisco updated its framework to address the new risks posed by agentic AI, as traditional security approaches focused on data and model integrity are insufficient for systems that can act and adapt independently.
03What new considerations does the updated framework include?+
The updated framework incorporates how agentic systems interact with their environment, make decisions, and potentially generate new risks without direct human intervention, moving beyond static data and algorithm security.
04What is the key challenge for responsible AI highlighted by agentic AI?+
The increasing autonomy of agentic AI systems demands a parallel evolution in oversight mechanisms, requiring organisations to move beyond static risk assessments to dynamic, continuous monitoring and 'controllability by design'.