AI risk management Archives

Agentic AI Systems: The Rise of Over-Autonomous Security Risks

1 July 2025 by Krishna

Artificial Intelligence (AI) is no longer just a tool—it’s becoming a decision-maker. With the emergence of Agentic AI Systems—AI with the ability to independently plan, act, and adapt across complex tasks—organisations are entering uncharted territory. While this autonomy promises operational efficiency, it also introduces over-autonomous risks that challenge traditional cybersecurity protocols.
For C-Suite executives and penetration testers alike, understanding the evolution of AI from a predictive model to a proactive actor is no longer optional—it’s imperative. The very qualities that make agentic systems powerful—initiative, goal-seeking behaviour, and environmental awareness—also make them vulnerable to sophisticated threats and capable of causing unintentional damage.

LLM06:2025 Excessive Agency — A Critical Vulnerability in the Age of LLM Autonomy

18 June 2025 by Krishna

The surge of Large Language Model (LLM)-driven applications has revolutionised how businesses interact with data, automate processes, and deliver enhanced user experiences. From autonomous customer service bots to intelligent data summarisation tools and generative co-pilots, LLMs are transforming enterprise workflows at an astonishing pace.
However, this rise has not come without significant risk. Among the top concerns identified in the OWASP Top 10 for LLM Applications v2.0, LLM06:2025 – Excessive Agency stands out as a particularly insidious and business-critical vulnerability. It affects systems where LLMs are entrusted not only with information retrieval or generation but with the ability to act on behalf of users — often through invoking external tools, plugins, or APIs.

LLM05:2025 – Improper Output Handling in LLM Applications: A Business Risk Executive Leaders Must Not Ignore

17 June 2025 by Krishna

At its core, Improper Output Handling refers to inadequate validation, sanitisation, and management of outputs generated by large language models before those outputs are passed downstream—whether to user interfaces, databases, APIs, third-party services, or even human recipients.

LLM04: Data and Model Poisoning – A C-Suite Imperative for AI Risk Mitigation

16 June 2025 by Krishna

At its core, data poisoning involves the deliberate manipulation of datasets used during the pre-training, fine-tuning, or embedding stages of an LLM’s lifecycle. The objective is often to introduce backdoors, degrade model performance, or inject bias—toxic, unethical, or otherwise damaging behaviour—into outputs.

Weak Model Provenance: Trust Without Proof

12 June 2025 by Krishna

Weak Model Provenance: Trust Without Proof A critical weakness in today’s AI model landscape is the lack of strong provenance mechanisms. While tools like Model Cards and accompanying documentation attempt to offer insight into a model’s architecture, training data, and intended use cases, they fall short of providing cryptographic or verifiable proof of the model’s …

Continue