Krish AI

AI-assisted dependency security analysis system that scans project dependencies, uses OSV for vulnerability information, uses an LLM to turn findings into a practical security brief, and requires explicit approval before patching.

• Python
• LangGraph
• Gemini 2.5 Flash
• OSV
• osv-scanner
• Rich

Dependency scanners dump vulnerability lists. Interpreting risk, deciding what to patch, and running upgrades still happens in a noisy, easy-to-get-wrong loop — especially if an agent is allowed to execute arbitrary shell commands.

Krish AI (KRAVEN) scans a target folder, maps package versions through OSV, asks Gemini 2.5 Flash to write a practical security brief, and only generates or runs package-manager patch commands after the operator approves. After patching, it re-runs the scan and summarizes remaining exposure.

• Single CLI entry (kraven.py) implementing a LangGraph-style phase workflow: scan → analyze → report → patch/verify.
• Primary scan path uses the osv-scanner binary when present; fallback reads pinned requirements, package-lock.json, and pip list from a venv.
• Rich-based operator console with phases, tables, and approval prompts.
• Remediation is limited to package-manager style commands inside the selected project directory.
• Safety layer blocks shell chaining, redirects, pipes, curl/wget, rm/del, and privilege-escalation paths.

Human approval is a control, not a UI flourish

No remediation runs without an explicit operator yes. The product is a triage assistant, not an autonomous patcher.

Constrain the shell

Even after approval, only package-manager commands are allowed. Arbitrary command composition is treated as a vulnerability of the tool itself.

Degrade scan coverage honestly

When osv-scanner is missing, fallback sources keep the tool usable. The README is explicit that native osv-scanner still has broader ecosystem coverage.

Evaluation data has not yet been benchmarked. There are no published false-positive rates, time-to-triage studies, or patch-success metrics.

• Not a replacement for an enterprise vulnerability-management platform.
• Patching is dependency/package-level only — not exploit remediation.
• Fallback scan quality depends on the metadata in the target project.
• Additional lockfile formats (poetry.lock, Pipfile.lock) are listed as roadmap, not current coverage.
• Analysis currently depends on a Gemini API key; this path is not fully local.