Who’s Watching the Agents? Security and Oversight for Autonomous AI

AI agents no longer just answer questions. They call tools, touch APIs, make decisions and produce more code, plans and text than any team can read line by line. That raises two questions every engineering team has to answer: how do we keep autonomous agents secure, and how do we keep up with reviewing what they produce?

This evening covers both sides. Google’s SAIF 2.0 states that agents need well-defined human controllers and that their actions and planning must be observable. We’ll first look at what that framework demands for securing agentic systems, and then at open-source tooling that makes human oversight practical, even when your agents produce far more than you can read.

THIS MEETUP will be in ENGLISH and HYBRID. | Zoom Link is visible for remote attendees after registration.

*** AGENDA ***

• 6:30-7:00 PM | Doors open. Grab some snacks & drinks.
• 7:00 PM | Welcome & Intro Start of the hybrid meetup
• 7:05 PM | SAIF 2.0 – Security Strategies for the Era of Autonomous AI Agents - Alexander Eimer (Senior Software Engineer at QAware)
• around 8:05 PM | Drowning in Diffs: Making Agent Output Readable Again - Niklas Selbach (Software Engineer at QAware)
• afterwards: More snacks & drinks, and a lot of time for networking with all attendees and speakers.

*** ABSTRACTS ***

SAIF 2.0 – Security Strategies for the Era of Autonomous AI Agents - Alexander Eimer (Senior Software Engineer at QAware)

The rise of large language models has made security a top priority for AI applications. However, while we are still in the process of securing static models, the next technological leap is already underway: the transition to autonomous AI agents. These systems don't just generate text; they also use tools, access APIs and make independent decisions.
In this talk, we analyse the Secure AI Framework (SAIF), developed by Google, and its evolution into SAIF 2.0. We will explain why conventional security approaches are ineffective with active AI systems and how robust agent infrastructures can be designed.
Key topics of the talk:

• SAIF Foundations: An overview of the six pillars for secure AI ecosystems.
• The leap to SAIF 2.0: Why agents need a new risk matrix.
• Technical Deep Dive: Analysis of new attack vectors, such as indirect prompt injection via RAG data sources, and security risks in agents' long-term memory.
• Governance in Practice: How the framework helps to develop “Secure by Design” AI applications and minimises risks through automated assessments.

This talk provides developers and security professionals with a clear roadmap for safely and responsibly unlocking the potential of AI agents in enterprises.

Drowning in Diffs: Making Agent Output Readable Again - Niklas Selbach (Software Engineer at QAware)

Coding agents can produce a plan, a twenty-file refactoring and a design document before you've finished your coffee. Generating content is no longer the bottleneck; understanding it is. Scrolling through walls of chat output or raw Markdown doesn't scale. When reading becomes a chore, human review of generated content quietly turns into rubber-stamping.

This talk argues that the answer isn't more discipline, but better formats. We climb a ladder of output formats, each one making agent output easier to take in:

• Clearer writing: prompting for plain, structured language that is quick to read and hard to misread.
• Diagrams: using agent skills to turn architectures, flows and data models into clean visuals.
• Interactive HTML: comparisons, collapsible sections and clickable artifacts you can explore instead of scroll through.
• Explainer videos: generated walkthroughs on any topic, and where they're already useful today.

But understanding is only half of a review. You also need to push back. With Plannotator and Lavish, you can mark the exact line, element or sentence that's wrong and send structured feedback straight back to the agent. We'll look at where this pays off in everyday work: reviewing specs and implementation plans in spec-driven development before any code is written, going through AI-generated code changes, challenging design documents and architecture proposals, and comparing alternative solutions side by side.

You'll leave with a practical, open-source toolkit for actually reading and steering your agents' output, without reading every line twice.

*****************
Please note that photos are taken during the meetup. If you do not want to appear in the photos, please contact the meetup team at the beginning of the event.
***
Während unserer Meetups werden Fotos aufgenommen. Falls du nicht auf den Fotos erscheinen möchtest, sprich bitte zu Beginn der Veranstaltung mit unserem Meetup Team.