The arrival of Generative AI, like ChatGPT or Gemini, a few years ago has made the world enter an endless journey of AI development, self-reliance, and supremacy. The next big shift is happening after Generative AI: we are moving towards agentic AI. It is an AI system that is capable of achieving a specific goal with limited or no human supervision. It is like a human acting on its own based on its logic and reasoning.

McKinsey Quarterly on Agentic AI

Based on McKinsey Quarterly, agentic AI presents both opportunities and risks for innovation, growth, and the acceleration of human civilization. In more than 60 generative AI use cases, such as customer service, software development, supply chain optimization, and compliance, agentic AI, which can reason, plan, act, and adapt without human oversight, is expected to release $2.6 trillion to $4.4 trillion in value annually.  Just 1% of surveyed firms believe that their AI adoption has reached maturity, despite the fact that many company leaders are rushing to embrace it because they believe it will completely transform the way they operate.

As a matter of fact, agentic AI may function without human supervision, it may present new internal dangers including interfering with business processes, jeopardizing private information, or undermining consumer confidence. The fact that 80% of the firms questioned reported experiencing dangerous AI agent behaviors, like inappropriate data exposure and unauthorized system access, is even more concerning.

Therefore, to address the risks and challenges associated with agentic AI and its agents, Chief Information Officers, Chief Risk Officers, Chief Information Security Officers, and Data Protection Officers can develop safety compliance frameworks. Agentic AI in the future will initiate actions, collaborate across silos, and make decisions affecting business. Therefore, it should be ensured that agentic AI does no harm, with trust as a foundational principle.

Emerging Risks in the Agentic Era

According to the McKinsey Quarterly report, risks associated with agentic AI include, but are not limited to:

• Chained vulnerabilities: For example, in which, due to a logic error, a credit data-processing agent misclassifies short-term debt as income, inflating the applicant’s financial profile.
• Cross-agent task escalation: For example, in which a compromised scheduling agent in a healthcare system requests patient records from a clinical-data agent, falsely escalating the task as coming from a licensed physician.
• Synthetic-identity risk: For example, in which an attacker forges the digital identity of a claims-processing agent and submits a synthetic request to access insurance claim histories.
• Untraceable data leakage: For example, in which an autonomous customer support agent shares transaction history with an external fraud-detection agent to resolve a query but also includes unnecessary personally identifiable information about the customer.
• Data corruption propagation: For example, in which, in the pharmaceutical industry, a data-labeling agent incorrectly tags a batch of clinical-trial results. This flawed data is then used by efficacy analysis and regulatory reporting agents, leading to distorted trial outcomes and potentially unsafe drug approval decisions.

However, while AI can provide abundant efficiency, such errors threaten to erode faith in the business processes and decisions that agentic systems are designed to automate. If the principles of safety and security are integrated prior to deployment, only then can agentic AI deliver its potential.

Guiding Principles for Agentic AI Security

Based on the McKinsey Quarterly report, there are guiding principles for agentic AI security. These are:

• Pre-deployment phase: Updating risk assessment techniques by creating AI governance; maintaining centralized AI portfolio management to prevent unchecked proliferation; building compliance with evolving regulations such as GDPR, ECOA, and local AI laws; upgrading AI policies for agentic systems, including IAM and third-party interactions; and guaranteed the capability of AI security personnel, threat modeling, and governance.
• Phase of deployment: Controlling access and behavior, guaranteed traceability, establishing contingency planning, and secure agent-to-agent interactions through authentication, logging, and permission management.

AI Bubble

These days in Silicon Valley, entrepreneurs are speculating whether the tech world is heading toward another bubble like the one seen in the 1990s. In the US, AI-related enterprises have accounted for 80% of the gains, with estimated global spending on AI expected to reach a whopping $1.5 trillion. Warnings regarding a potential AI bubble have come from the Bank of England, the IMF, as well as JP Morgan.

Source: A Historical Overview of AI Winter Cycles

AI Winter and AI Summer are terms used to describe the historical cycles of AI development.  AI Summer is a time of increased interest and investment in AI, while AI Winter is a time of decreased interest and investment.  The first AI Summer was through the period of 1950s and 60s, 1970s saw an AI Winter, the 1980s saw another AI Summer, and the late 1980s and early 1990s saw yet another AI Winter. With the growth of machine learning after 2000, AI entered another summer. Deep learning then accelerated further in the 2010s, and generative and agentic AI are currently enjoying an AI Summer in the 2020s.

In conclusion

Technologically and economically, we are moving on a linear path. Therefore, AI is unavoidable but equally requires safety and resilience frameworks. To avoid catastrophes, organizations must strike a balance between innovation and risk management through proactive controls and governance.