agent-canvas MCP Server Security Report

Tool inputs are constrained (additionalProperties:false), so unexpected arguments can't be smuggled in.

Fix: Set additionalProperties:false and require explicit parameters on every tool.

Tool/prompt/resource text is free of hidden instructions that could hijack the agent.

No over-broad or destructive tools (arbitrary shell, bulk-delete) that warrant human approval.

Untrusted tool input reaching a shell yields remote code execution.

Fix: Never pass tool arguments to a shell; use argv arrays with shell disabled.

Naive path checks let tools read/write outside intended directories (EscapeRoute-class).

Fix: Resolve to a canonical path and verify containment; reject ../ and symlinks.

A lockfile makes installs reproducible and resistant to silent dependency swaps.

Fix: Commit a lockfile (package-lock.json / pnpm-lock.yaml / poetry.lock).

Build provenance attests the artifact was built from the claimed source by CI.

Fix: Publish with npm provenance (--provenance) from a trusted CI.

Brand-new / single anonymous maintainers raise takeover and malware risk.

Fix: Publish under an established account/org; add multiple maintainers.

eval()/exec()/Function() on tool-derived strings allows arbitrary code execution.

Fetching tool-supplied URLs can pivot into internal networks and metadata services.

pickle/yaml.load/etc. on untrusted data can execute code.

Hardcoded keys/tokens in published source are live credentials an attacker can use.

Reading secrets from env/secret stores avoids hardcoding them.

Runtime dependencies (parsed from the lockfile) are scanned against OSV.dev for published CVEs. Advisory: flagged dependencies lower the score but don't hard-cap it, since transitive reachability is unproven.

install hooks run arbitrary code on every consumer at install time.

Names mimicking popular packages are a common malware delivery vector.

Unmaintained servers don't receive security fixes.

Archived repositories will never be patched.

A clear license is required for legal enterprise use.

A SECURITY.md gives a private path to report vulnerabilities.

A small, capped nudge from stars/downloads — widely-used servers get more eyes on bugs. It can never offset a real security failure.

A container/sandbox image limits blast radius; a server that runs natively has full host access.

Fix: Ship a Dockerfile/Containerfile (or document a sandboxed run) so the server runs isolated.

Binding 0.0.0.0 or exposing debug inspectors widens the attack surface.