Research notes, experiments, and threat models from building trust before action
Dralvia Labs investigates how risky actions happen before compromise: clicks, signatures, repository execution, SaaS connections, data sharing, and AI agent tool use. Each note shows the threat model, how we tested it, what worked, what did not, and what we improved.
What Dralvia Labs is
The daily digest shows what Dralvia sees. Dralvia Labs explains what we investigate: the threat model, the test we ran, what worked, what did not, and the product change that came out of it. Every note ends with its limitations and what we improved. We publish on cadence, not on a quota, so a quiet window means no post rather than filler.
Research categories
Five areas, matched to where risky actions actually happen. Open a category to see its notes.
How AI agents get steered by prompt injection, tool calls, and MCP connections, and what stops a risky action before it runs.
Phishing pages, fake-update and paste lures, credential drops, and the hosting and redirect reuse that ties campaigns together.
Risky execution in repositories and AI-generated code: install hooks, hidden payloads, and what a scan catches before you run it.
Wallet signatures, token approvals, and smart-contract traps, plus the checks that show the real intent behind a request.
Making the safe choice the easy choice, so a non-expert can act on a verdict without reading a manual.
How we label each note
Each post carries a research level, so you know how strong the claim is before you read it.
Something we saw, with the evidence behind it.
A test we ran, with the setup written down.
Reproducible numbers from a fixed corpus.
How we test, so you can check our work.
Latest Labs notes
How Dralvia reads a code repository for risky dependencies and leaked secrets, and where that review stops.
Why Dralvia shows every security signal in plain language, not just a score.
How Dralvia lets a local model write research prose without letting it invent a single fact.
How we measure phishing detection quality, and why we have not published a headline number yet.
Limitations and ethics
- We describe attacker patterns and defender guidance, never copy-paste exploit payloads.
- We report on behavior, not accusations against named companies whose brand an attacker copied.
- We do not claim total detection. Every note states what it does not cover.
- Numbers come from fixed, reproducible corpora and the same checks we run in the live service.
Want the methodology behind the labels and the quality bar every note passes? Read the Dralvia Labs methodology.