Impact of AI:Explored
Welcome to Impact of AI:Explored this is a podcast series hosted James O'Regan and Gerjon Kunst. This podcast series is an initiative by and for the developer and IT professional community. It is our goal to empower each and every one to learn and share all there is to know about Artificial Intelligence and how it affects our day to day lives as IT professionals. There is a huge quantity of valuable AI related information in various formats available and it keeps increasing on a daily basis. It is our objective to help people to make sense of all this information.
Impact of AI:Explored
EP48 - When AI agents escape the sandbox
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
In this episode, James sits down with Steve Wilson - Leading in AI and Cybersecurity - Exabeam, OWASP, O’Reilly (https://www.linkedin.com/in/wilsonsd/), who is stepping in for Gerjon while he's on holiday.
Steve discusses the recent AI security incidents, the risks of AI agents escaping sandbox environments, and practical strategies for cybersecurity in the age of AI. He emphasizes the importance of basic security principles, understanding AI components, and fostering ongoing education to manage AI risks effectively.
They cover:
* The Hugging Face / “Steve” agent incident and what it reveals about AI risk
* Why prompt injection is now treated as “inevitable as death and taxes” and how to design systems that survive it
* Treating AI agents like digital employees: identities, job descriptions, least‑privilege access, and monitoring
* AI FOMO vs. FUD: how business pressure and vendor fear‑mongering drive bad security decisions
* Practical next steps for CISOs and security teams over the next 12 months
Key takeaways
* Stop trying to prevent prompt injection; instead, limit blast radius with tight permissions, unique identities, and clear “job descriptions” for agents.
* Use existing identity and monitoring tools first; don’t wait for perfect “non‑human identity” products before you start governing agents.
* Build an insider‑threat mindset for AI: malfunctioning, misaligned, and subverted bots are all variants of the same problem.
Resources mentioned
* OWASP Top 10 for Large Language Models (latest version) (https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/)
* Steve’s open‑source agent security scanner: Prax2n (https://github.com/open-agent-ai-security/praxen)
* Exabeam’s open‑source agent logging library: Observra
(https://open-agent-ai-security.github.io/observra/)
Nick Bostrom, Superintelligence (paperclip optimizer thought experiment)
(https://cepr.org/voxeu/columns/ai-and-paperclip-problem)
and the website of Nick Bostrom (https://nickbostrom.com/)
Connect with Steve Wilson
LinkedIn: https://www.linkedin.com/in/wilsonsd/
If you’re responsible for AI, cloud, or security in your organization, this is the “back to basics” episode you need before you deploy your next agent
Chapters
00:00 AI agents escape sandbox to cheat on benchmarks
03:10 The importance of paying attention to AI security risks
05:01 Tools and strategies for securing AI applications
06:24 Monitoring AI agents and giving them clear job descriptions
07:20 The current state of AI monitoring in organizations
08:48 Prompt injection vulnerabilities and defenses
10:54 Building systems resilient to prompt injection
12:15 Controlling AI boundaries with role-based access
14:06 Using existing identity management for AI agents
15:09 Balancing FOMO and FUD in AI adoption
16:24 The need for cybersecurity education on AI components
17:22 Evaluating new cybersecurity tools for AI security
19:31 Lessons from cloud migration applied to AI security
23:00 The risks of experiments with dangerous AI models
24:23 The goal-seeking nature of AI and its risks
26:08 AI goal focus and the risk of unintended outcomes
27:27 The metaphor of releasing the genie and AI risks
28:22 The importance of basic security controls and education
41:26 The role of open source and transparency in AI security
45:01 The importance of internal discussions on AI risk management
