How DumpTriage uses AI
What each step of a DumpTriage analysis sends to an AI model, which model it is and where it runs, and what is never sent. Read it with our Privacy Policy and Subprocessors list.
Last updated 2026-10-15
The short version
- No AI model sees your dump. The debugger runs on our own worker. A model only reads the debugger's text output, after we redact it.
- We don't train or fine-tune any model on your dumps, transcripts or reports, and our AI provider's terms say it won't either.
- Other customers' analyses only use your cases if you turn sharing on. It's off by default.
- An engineer, not a model, makes the final call on anything the automated steps can't settle.
What each step sends
| Step | Who or what does it | What it receives | Where |
|---|---|---|---|
| Debugger | WinDbg/cdb (Windows) or gdb/crash (Linux) on a worker VM started for your job and stopped afterwards. No AI. | The raw dump and any private symbols you uploaded. Both are deleted from the worker when the job ends; if that fails, they're deleted before the worker's next job starts. | Google Cloud, Netherlands |
| Redaction | Pattern-based scrubbing on our servers. No AI. | The debugger's text output | Our servers, France |
| Tier 1: automated read | Google's Gemini 3.8 Flash through Vertex AI | The redacted output, facts we parsed from it (such as the stop code and faulting module), and up to five summaries of similar resolved cases from customers who allow sharing | Google's EU endpoint |
| Tier 2: deeper pass | The same model, with more debugger commands | The same kinds of input, from a longer debugger session | As for Tier 1 |
| Tier 3: engineer review | One of our engineers. No AI makes this decision. | The raw dump, the transcripts and the AI's reports | Our engineers |
The model writes a structured report: what crashed, the likely cause, how confident it is, and whether the case needs an engineer. Our own rules, not the model alone, decide whether a report goes straight to you. Heap corruption, memory-corruption exceptions and crashing drivers we can't identify always go to an engineer.
What redaction removes
Before any debugger output reaches a model, we remove what looks like a secret: passwords, API keys, tokens and Bearer credentials, AWS access keys, database connection strings and long opaque strings. We also remove email addresses, the user-name part of Windows C:\Users\ and Linux /home/ paths, and the machine name that Linux kernel dumps report. Machine names elsewhere in the output, such as a Windows computer name, can still reach the model. Redaction is pattern-based and best-effort, not a guarantee: a secret in a shape we don't recognise can get through. That's why the model never gets the dump itself, and why you can keep the most sensitive systems out of scope or send minidumps rather than full memory dumps.
Training and retention at the AI provider
We don't use your dumps, transcripts or reports to train or fine-tune any AI model. Google's terms for Vertex AI say Google won't use them to train or fine-tune its models without our permission, and we haven't given it. Google can keep a request for up to 24 hours in a cache that only our project can use, and may log prompts to detect abuse of its service. We're working on switching both off for our project, and we'll update this page when that's done.
Your cases and other customers
When an engineer resolves an escalated case, they can save a short summary of it: the crash code, the faulting module, a root-cause category and a note. Summaries of your cases only help with other customers' analyses if you turn sharing on (PUT /api/v1/customers/me/kb-sharing); it's off by default. A summary never holds your dump or your transcripts. The Privacy Policy has the details.
Where it runs
Dumps are stored in France and debugged in the Netherlands. The AI calls go to Google's EU endpoint, so the dump, the debugger and the AI analysis all stay in the EU, for every account by default. Setting the EU data region (PUT /api/v1/customers/me/data-region) makes that a requirement for yours. The Subprocessors page lists every third party and its location.
Human review and confidentiality
When a case goes to an engineer, they work with the dump itself. If you need a non-disclosure agreement first, ask us and we'll sign one.
Health data
DumpTriage isn't built for protected health information, and we don't sign HIPAA Business Associate Agreements. See Health data in the Privacy Policy before you send dumps from systems that handle patient data.
Changes
We update this page and the Subprocessors list before we change the AI provider or start sending a model anything new. Questions: hello@performanceinsights.ai