Why Gateways Break AI Operations (And How to Actually Secure Agents) | Hanah Darley, Geordie AI

Why Gateways Break AI Operations (And How to Actually Secure Agents) | Hanah Darley, Geordie AI

by AI Security Podcast · View the channel on PSI

Applied Sciences & Engineering AI — News & Applications UYComputer scienceUYQArtificial intelligence (AI)
🎙 AI Security Podcast 👥 25K 📅 September 17, 2026 ⏱ 53 min 👁 2K 📄 expert opinion 🧭 2026-09-18
Available in: English (current) Français

Keywords

AI securityagentic AIgatewayscontext engineeringhuman-in-the-loop

Summary

This episode of the AI Security Podcast features Hanah Darley, Chief AI Officer at Geordie AI, discussing the challenges of securing agentic AI systems. The conversation begins by questioning the term ‘AI security’ as too broad, suggesting instead a focus on specific applications like LLMs or agents. Darley emphasizes that traditional gateways and firewalls introduce latency and fail to handle the dynamic, goal-pursuing nature of agents. She argues that logs are insufficient for understanding agent behavior, which requires continuous classification and behavioral analysis. The discussion covers the full agent lifecycle, from creation to live operations, and highlights the need for context engineering to influence agent decisions in real time. Darley also critiques over-reliance on human-in-the-loop, noting that humans often perform worse than AI in objective tests. She proposes a maturity model for AI security programs, starting with explicit red lines and advancing to autonomous decision-making. The episode concludes with a discussion of Geordie AI’s agent-out deployment approach and a rejection of the ‘AGI pill’ in favor of pragmatic security measures.

170 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of this episode lies in its practical, experience-backed insights into AI agent security, a rapidly evolving field. Darley challenges common assumptions, such as the effectiveness of gateways and kill switches, offering a nuanced perspective that distinguishes between AI native and AI first organizations. Her argumentation is coherent and grounded in her background in government intelligence and cybersecurity, though it relies heavily on anecdotal evidence and personal observations rather than formal studies. The discussion of cognitive biases affecting security professionals is particularly insightful, as it explains why legacy approaches fail with non-deterministic AI systems. However, the lack of concrete data or case studies weakens the empirical foundation of the claims.

Scientific Rigor, Source Quality, Title Accuracy

The rigor is moderate: the conversation is expert-driven but does not cite specific research or external sources. The only sources provided are the podcast’s own website, newsletter, and LinkedIn page, which offer no further reference material. The title is well-matched to the content, as the host and guest spend significant time explaining why gateways break AI operations and propose alternatives like context engineering. A notable omission is the absence of references to incidents like the OpenAI หรือ Anthropic agent escapes, which are mentioned but not detailed. Given the exploratory nature of the podcast, the lack of citations is understandable, but it limits the scientific authority of the claims. The public’s engagement is low (5 likes, 1,511 views), suggesting limited reach, but no comment analysis is available to assess audience reception.

256 words

Title / Content Match

The title accurately reflects the core themes: the shortcomings of gateways in AI operations and alternative approaches to securing agents.

Quality & Reliability

6/10

The discussion is based on the guest's professional experience rather than peer-reviewed research. It offers practical insights but lacks formal citations or empirical evidence, resulting in moderate reliability.

Chapters

Cited Sources

Contribution & Novelties

This episode contributes original perspectives on AI agent security, particularly the critique of gateways as a latency-inducing, binary enforcement mechanism and the introduction of ‘context engineering’ as a dynamic method to influence agent behavior. The maturity model proposed offers a practical roadmap for enterprises. The discussion also highlights cognitive biases in cybersecurity professionals that hinder adoption of new approaches. For further exploration, the following resources are relevant:

110 words

Radar Profile

The radar profile indicates a high quantity of information (8) and technical level (7), but lower reliability (6) and moderate quality (7). This suggests a rich, technical discussion that is opinion-driven rather than empirically rigorous, fitting an expert podcast format.

Reliability 6/10