Power Strip Security

Prompt-injection scanner for content your agent is about to read. Returns a risk score (0-1), a level (low/medium/high), findings and the cleaned text wrapped as untrusted data (hidden text, invisible characters, data-exfiltration links, instruction-override phrases in 13 languages). Deterministic. Content is scanned in memory and never stored. Not charged if it fails. Also: /v1/read $0.005, /v1/s

Safe web reader for agents: give it a URL, get the page as clean markdown with hidden text, invisible characters and data-exfiltration links removed, a prompt-injection risk report, and the content wrapped as untrusted data. Respects robots.txt. Pages it cannot read (blocked, needs JavaScript, too large) are not charged.

MCP tool-poisoning audit: check an MCP server before your agent connects. Send its URL or a pasted tools/list; get findings on hidden instructions, cross-tool shadowing, secret-file access and exfiltration parameters, plus a fingerprint per tool to catch rug pulls. Read-only, never calls a tool. Unreachable servers are not charged. MCP server: npx -y power-strip-security-mcp.

Agent skill audit: check a SKILL.md folder (scripts, references) or AGENTS.md, CLAUDE.md, .cursorrules before an agent loads it. Findings by file and line: prompt injection, curl | bash, encoded commands, credential theft, exfiltration, hardcoded keys, persistence, spec errors. Paste the files or one URL. Nothing runs, nothing is stored. Cheap enough for every PR.