Cryptographic audit trails for AI workflow compliance

[jagmarques]

Flyte orchestrates AI workflows with data, models, and compute. In regulated environments, teams need more than execution logs - they need provable, tamper-proof records of what ran, with what data, and what the outputs were.

asqav adds cryptographic signatures to workflow steps. Each task execution gets a signed record - useful for proving to auditors that a model training run or data processing step happened exactly as reported.

Could integrate as a Flyte plugin that wraps task execution with signed audit records, especially useful for ML pipelines in finance and healthcare.

[runllm]

It seems likely that Flyte already provides strong data lineage, versioning, and audit trails for workflow executions, including capturing inputs, outputs, and metadata for each task and workflow run. However, Flyte's built-in audit trail is not explicitly described as cryptographically tamper-proof or signed; it focuses on immutability, reproducibility, and traceability by storing execution artifacts and metadata in object storage and catalog systems, with each execution being versioned and immutable. This provides a robust audit trail, but not cryptographic guarantees out of the box. See Using Flyte to Power MLOps at Striveworks, Build Indestructible Pipelines With Flyte, and Declarative Infrastructure: The Power of Unmatched Reliability.

Integrating a system like asqav as a Flyte plugin to add cryptographic signatures to each task execution is not described in the current data sources, but Flyte's plugin architecture is designed for extensibility, including custom task plugins and backend plugins. This suggests that such an integration is feasible, especially for regulated environments needing cryptographically signed audit trails, but there is no official or community-supported plugin for this purpose documented yet. See Extending Flyte.

Would you like more detail on Flyte's existing audit and lineage features, or on how to build a custom plugin for cryptographic audit trails?

Sources:

• Extending Flyte
• Using Flyte to Power MLOps at Striveworks
• Build Indestructible Pipelines With Flyte
• Declarative Infrastructure: The Power of Unmatched Reliability

---

This is an AI-generated response and your feedback is appreciated! Please leave a 👍 if this is helpful and 👎 if it is not.

Mention @RunLLM to ask follow-up questions. See runllm.com for more!

[muskan-creates352]

Hi,

This idea looks really interesting, especially for compliance use cases.

I’m interested in contributing to this. Would it be okay to start by exploring a minimal prototype or plugin approach for adding signed audit records to task executions?

Any guidance on where this could fit best in Flyte’s architecture would be really helpful.

Thanks!

[jagmarques]

Hi @muskan-creates352, yes please go for it.

Cleanest fit is a flytekit decorator wrapping @task that emits signed receipts at the queued/started/finished hooks. The Asqav SDK already does the signing (pip install asqav, free tier is enough for a spike): https://github.com/jagmarques/asqav-sdk

Happy to review a draft PR.

[falloficaruss]

#take

[falloficaruss]

@jagmarques Since no PR was raised by @muskan-creates352 , I have implemented the asqav audit plugin
You may review it

[jagmarques]

@falloficaruss thank you for picking this up and shipping a full implementation. The PR is at flyteorg/flytekit#3447. We left a detailed review there with a couple of items worth addressing before merge. Happy to help land it.