The next leg of DataJourneyHQ is about going deeper into practical adversarial testing for AI agents and turning what we learn into guidelines other builders can use. We are glad to explore that work through the NVIDIA Inception program.
For us, the program opens a closer path to technical resources, training and a wider startup ecosystem as we learn, test our ideas and keep doing careful, useful work.
Connecting the dots
Our time with GitHub’s Secure Open Source Fund changed how we think about security. Tools mattered, but the lasting lesson was to turn security into a habit: make permissions visible, checks repeatable and ownership clear before a project becomes large enough to hide its risks.
AI agents widen that boundary. An agent does not only produce an answer. It can choose a tool, pass data, retry a failed step and sometimes trigger an action in another system.
This is where the DataJourneyHQ view of data matters. Data is not cargo moving from one box to another. It passes through people, policies, tools and models, picking up context and consequence along the way. We want DJHQ to sit at the centre of that flow and help make the whole path understandable.
The AI layer brings useful non-determinism: the ability to interpret, adapt and find a path through a problem. Around it, we still need deterministic boundaries such as schemas, permissions, budgets, stop conditions, approvals and audit trails. One gives the system room to work. The other gives people a reason to trust it.
What we want to build next
We want to make adversarial testing for agents more practical, especially for small teams that do not have a dedicated AI security group.
That means working towards:
- Test cases that follow the full agent journey from input to tool call to action
- Simple ways to probe prompt injection, poisoned tool responses, excessive permissions, unsafe retries and secret leakage
- Evaluations that inspect tool use and visible intermediate steps, not only the final answer
- Human checkpoints that appear where an action becomes consequential
- Open exercises and lessons that builders can use through DJHQ Academy
Using NVIDIA tools with care
As we begin, we want to explore where NVIDIA NeMo Guardrails and the NVIDIA NeMo Agent Toolkit can strengthen the work.
NeMo Guardrails can place programmable checks around model inputs, outputs and tool boundaries. The NeMo Agent Toolkit can help us run agent workflows against test data and examine how they use tools, not only what they say at the end.
Neither is a shortcut to a trustworthy agent. Our plan is to start small: build prototypes, keep the test cases versioned, record where a guardrail holds or fails and bring the useful patterns back to the community.
Closing the loop
We do not want to write these guidelines in isolation. The most useful edge cases come from open-source maintainers, security practitioners, product teams and builders who understand both the promise of agents and the systems they touch.
That is the loop we want DJHQ to hold together: learn from real systems, test the weak points, turn the evidence into practical guidance and teach it through the Academy so the next builder can begin with stronger defaults.
This is a small beginning for the next phase. The work ahead is still the same kind of work we believe in: one threat model, one test and one better boundary at a time.