The record a security review can sign off
Grounding an agent in a shared store only helps if the store itself is defensible. becos's answer to that is provenance, structural isolation, and injection-safety — the checklist a review runs against.
What's in the record
- Prompts and the agent's text replies
- Commit id, branch, author, changed-file counts
- File paths
- Distilled Markdown entity pages + a dated ledger
What never enters it
- Diffs / patch content
tool_use,tool_result,thinkingblocks- Error strings (only the exception type)
- Anything from another tenant — dropped before the model
Why the record isn't a new attack surface
Provenance you can byte-check
Every fact resolves to a stored event via resolve_event — verbatim, no model in the loop. Grounded answers, not a second hallucination surface.
Structural isolation
Pages live under a normalised tenant path; results are filtered by tenant prefix before content reaches a model. An earlier shared-tenant design was explicitly retired as a hazard.
Injection-safe by construction
The record is treated as untrusted. Its content can't express slug / path / tool fields; nothing that writes it holds an action channel; failures are logged content-free.
Where it lines up with the frameworks
Framing for a controls conversation — not a certification claim. becos is currently a Phase 1 deployment; per-user OAuth is in progress.
event_id keys link a grounded claim to identity, timestamp and resource.Get the security narrative early
So the platform team's deal clears review instead of stalling in it.
Talk to us