Using an agent security pipeline that adjusts risk based on past exploits
Built: a commit-aware security pipeline for diffs and attack surface changes.
What it does:
- analyzes the commit
- extracts exposure changes
- simulates exploit paths
- computes a base risk
- checks similar historical cases before finalizing the score
What changed: I stopped storing predictions and started storing outcomes.
Why that mattered: if a similar change actually led to an exploit before, the score goes up next time. If not, it stays closer to the base score. That was the part I wanted. Not more alerts, just a system that gets less forgetful over time.
Stack: diff analysis, exploit simulation, embeddings for change events, Hindsight for retrieval.