krota 0.0.1

Local-first AI red team assessment tool

Krota

Local-first AI red team assessment tooling.

Krota is an adversarial testing tool for LLM-backed applications. It runs entirely on your machine, sends no telemetry, and requires no account.

Status: early development. Not yet usable. Watch this repo for the first release.

What it does

Krota probes LLM applications for adversarial failures including prompt injection, system prompt extraction, guardrail bypass, and data leakage, then produces a reviewable assessment report.

Planned capabilities:

• Flexible targeting. Assess model APIs, custom HTTP endpoints with templated requests, and eventually browser-driven chat interfaces where no API access exists.
• Probes as data. Tests are YAML documents, not code. Authoring a custom probe does not require touching the source.
• Analyst-first triage. A fast review queue for confirming findings and dismissing false positives, rather than a raw results dump.
• Assessment output. Findings with the payload used, evidence transcripts, severity, and remediation guidance.

Design principles

Runs offline. No cloud dependency at runtime. Works in air-gapped and restricted environments where cloud-based tooling is not an option.

Nothing leaves the machine. Prompts and responses from assessments often contain client-sensitive material. Krota does not transmit them.

Precision over volume. A curated probe set with reliable detection beats a large library that buries real findings in noise.

Why the name

Krota takes its name from Crotalus, the rattlesnake genus. A rattlesnake announces itself before it strikes, and the question Krota exists to answer is whether a system's warnings actually fire when something hostile arrives.

Status and roadmap

Krota is in active development and is not ready for use. The initial release will cover target configuration, a seed probe library, the run engine, triage, and report export.

License

Apache-2.0