Skip to content

    Type a page, a command such as keryn scan, or a rule id.

    What Keryn is

    View as Markdown

    Keryn maps the trust boundaries in software that uses AI and shows which edges an attacker can cross, with the proof for each.

    An edge is one thing reaching another: user input into a prompt, model output into a shell, a route into an agent’s tool, a workflow into an outside service. A trust boundary is an edge where authority changes. Why the edges defines the terms, and the edges section is the manual.

    Point Keryn at a repository and it reports:

    • Trust boundaries. Where input becomes instruction, output becomes action, or an agent reaches a tool, each with the file and line.
    • Paths. Data flow from a source to a sink, within a file and across files, kept as a trace.
    • Proof on every finding. From a pattern match up to a reproduction. See the proof ladder; the glossary defines the decision, analysis, severity and evidence words.
    • The AI-BOM. The models, provider SDKs, agent frameworks, MCP servers, vector stores, prompts and tools the edges attach to, as CycloneDX 1.6.
    • Controls. The findings mapped, by reference, to OWASP LLM and Agentic, MITRE ATLAS, NIST AI RMF, ISO/IEC 42001, the EU AI Act, NIST SP 800-53 and ISO/IEC 27001, with a posture scorecard and hand-off reports. See controls and posture.
    • Code, dependencies and binaries. Two source engines: the pattern engine matches rules, and Deep follows data across files and confirms or questions the pattern findings (static analysis). Open-source dependencies (SCA), committed secrets and agent skill files, compiled helpers and strings dumps (binary analysis), and agent workflow exports (workflows).
    • AI attribution. Which AI coding tools a repository is set up for and which commits declare them, with a confidence and the evidence for each. See AI attribution, the keryn attribution command and the POST /api/v1/ai-attribution endpoint of the Cloud API.
    • Trust you can model and decide on. The Trust Canvas draws the edges of an agent, a workflow or a whole organization and finds the chains across them; release decisions record who accepted which risk.
    • A policy gate and a coverage statement. A short YAML file says which providers and components are allowed. Every report says how many rules ran, and exit code 3 flags an incomplete run. See policies.

    Keryn runs where your code already is: Keryn CLI and CI, editors, Keryn Web and the browser extensions, and Keryn Desktop. Apps and integrations lists every surface, what it is for and whether it is released yet, and coverage lists what each kind of scan checks. Every surface runs the same engine, compiled to WebAssembly, on your machine. Nothing leaves your machine unless you press a button that names the destination (privacy). The one hosted part is the optional Keryn Cloud, which scans an address you type from Keryn’s servers and never receives your code.

    The product map lists every Keryn product with its status on each surface. The capability map describes the current browser scanner, host-specific integrations and the planned ASPM, continuous validation, runtime, supply-chain, API, cloud and AI/agent security scope. Check it alongside evaluation results; an engine benchmark is not a release manifest.

    Keryn does not replace a general static analyser and does not prove that code is safe. A scan reads files and never runs them; the dynamic tests (DAST and AI security tests, in the next release) are separate commands that send requests only to a target you name. Its detection limits are measured and published in what Keryn does not do. Inventory is the strong part: every entry points at a line you can check.

    Keryn is in pre-release. The install page says which builds exist today and how to get them.