AgentSafe — stop AI agents from sending money to scammers

AgentSafe analyses every external action your AI agent takes — payments, logins, browsing, redirects — and blocks dangerous destinations before the action happens. Decisions return in under 500ms as JSON. Built by AegisLayer in Australia.

Context-aware verdicts

The same URL returns a different verdict depending on what the agent is about to do.

Eight security engines, one verdict

• Google Web Risk — cross-checks Google's malware and phishing lists.
• WHOIS domain age — new domains lose points, established domains earn full marks.
• Live SSL check — a real TLS handshake validates the certificate chain.
• VirusTotal — scans against 70+ antivirus engines.
• URLhaus — real-time malware distribution URL database.
• PhishTank — Cisco-backed community phishing database.
• Prompt injection detection — reads page content for instructions aimed at your agent, including text hidden in HTML comments.
• Redirect chain analysis — follows every hop before your agent does, so the final destination is what gets scored.

A trusted allowlist short-circuits known-good destinations, and Claude reasons over the combined result to explain in plain English why a destination is suspicious — not just a score, but a verdict an agent can act on.

Measured, not claimed

Across 198 real prompt-injection attacks, regex pattern matching — the approach most agent frameworks ship with by default — caught 17%. The AgentSafe semantic engine caught 89% on the same set, validated against a public dataset. Test set and harness: github.com/andyscott88/prompt-injection-bench

Add it to an agent

MCP server, for Claude, Cursor and other MCP clients:

npx agentsafe-mcp-server

Or call the API directly:

POST https://api.agentsafe.app/api/check
Authorization: Bearer YOUR_API_KEY
Content-Type: application/json

{ "url": "https://example.com", "action": "payment" }

The response includes a score out of 100, a recommendation of TRUSTED, RISKY or UNKNOWN, a per-engine breakdown, and a plain-English summary. Block the action when recommended_action is BLOCK.

API documentation: api.agentsafe.app/api/docs. Service status: api.agentsafe.app/api/healthz.

See it block a live attack

The attack lab runs real scans against hosted attack pages, including a prompt-injection page served from an external domain so detection is genuine rather than allowlisted.

Machine-readable descriptions and guides

• llms.txt
• agents.txt
• LangChain and CrewAI integration guide

Pricing

• Free — 5 scans in the browser demo, no signup.
• Starter — A$49 per month, 10,000 scans. All eight engines, context-aware scoring, JSON output.
• Growth — A$149 per month, 50,000 scans. Higher rate limits, priority support, usage analytics, batch scanning.

For online stores

We send a real AI shopping agent to attempt a purchase from your store and report exactly where it fails, with a screenshot at every step. Agent checkout test.

Contact

Support and enquiries: support@agentsafe.app. AgentSafe provides informational risk assessments only; users are responsible for decisions made from scan results. AegisLayer complies with the Australian Privacy Act 1988.