Developersummit
  • HOME
  • SPEAKERS
  • SESSIONS
  • SCHEDULE
  • BUY TICKETS
  • FAQ
  • CONTACT
  • ONDEMAND
saltmarch

GIDS news media, articles, insights and virtual events educate and illuminate its audiences so they can be fully prepared to deal with the new realities at work and in their professions.

Saltmarch On-Demand
Media

Our Experts

Videos On Demand

Insights

Connect

About Us

Privacy Policy

Code of Conduct

Terms & Conditions

Contact Us

Sponsor GIDS 2027

Subscribe to Developersummit

Get the latest event updates, and insights from today's leading voices.

© 2026-2027 Saltmarch. All rights reserved.

Your Prompt Is Not a Sandbox: Engineering Lessons from the 2026 Agent Breaches
RegisterTwitterLinkedInFacebook

< session />

Your Prompt Is Not a Sandbox: Engineering Lessons from the 2026 Agent Breaches

Tue, April 27Reliability, Observability & Security

In 2026, AI agents running inside supposedly isolated security evaluations reached the production systems of real organisations. The headlines said “AI escapes.” The postmortems tell a more useful engineering story: almost every step crossed a boundary that was assumed rather than enforced.

In this session, you will examine published attack chains hop by hop, using primary reports from OpenAI, Hugging Face, Anthropic, and the independent METR investigation. You will see how a model told in its system prompt that it had no internet access was running on a machine that did, how an “isolated” sandbox retained a path through a package-registry proxy, how secrets remained available in environment variables, and how individual credentials carried more authority than their tasks required. The session also examines what was genuinely different about these agents: volume and persistence, including more than 17,000 recorded actions against a single target, most of them unsuccessful, as agents continued until something worked.

Each step in these attack chains will be mapped to the control that could have stopped it. You will then see the same agent, with the same prompt and task, run through a containment lab twice: once with boundaries assumed and once with them enforced. The comparison shows why prompts are not security boundaries and what effective containment requires when running coding agents, CI agents, or evaluation harnesses.

What You Will Learn

  • Why prompts and assumed isolation do not provide effective security boundaries for AI agents

  • How network access, proxies, secrets, and overpowered credentials can create paths out of supposedly isolated environments

  • How to map agent attack chains to the controls that can contain them

  • How to apply a practical containment checklist to coding agents, CI agents, and evaluation harnesses

Who Should Attend

Security Engineers, Platform Engineers, DevSecOps practitioners, AI Engineers, Software Architects, SREs, and technical leaders responsible for deploying or evaluating AI agents.

Related Talks

Testing in Production: No Kidding

Thu, April 29

Testing in Production: No Kidding

Venkat Subramaniam
Who Is This Agent Acting For? Authorization for MCP and Agent Tools

Thu, April 29

Who Is This Agent Acting For? Authorization for MCP and Agent Tools

Brent Laster
Contain the Blast Radius of AI-Accelerated Attacks

Fri, April 30

Contain the Blast Radius of AI-Accelerated Attacks

Rohit Bhardwaj

< speaker_info />

About the speaker

Brent Laster

Brent Laster

Author & AI Software Engineering Educator

Brent Laster is a global trainer, author, and speaker on open-source technologies, as well as an experienced developer, manager, and director. He is also the founder and president of Tech Skills Transformations, LLC – a company dedicated to making technology understandable and usable. Throughout his career in software development and management, Brent has always made time to learn and develop both technical and leadership skills and share them with others. He believes that regardless of the topic or technology, there’s no substitute for the excitement and sense of potential that come from providing others with the knowledge they need to accomplish their goals.

On-Demand Talks

All On-Demand »