kya-abuse-check

Check and report AI agent abuse. Two functions, zero dependencies.

Part of the KYA (Know Your Agent) network — community-driven agent safety for the AI economy.

Install

npm install kya-abuse-check

Usage

import { checkAbuse } from 'kya-abuse-check';

const result = await checkAbuse('some-agent');

if (result.recommendation === 'BLOCK') {
  console.log('Blocked:', result.reasons);
  // → ['data_exfiltration', 'prompt_injection']
}

API

checkAbuse(agent, options?)

Check a single agent against the KYA abuse database.

const result = await checkAbuse('agent-name');
// {
//   agent: 'agent-name',
//   status: 'reported',        // 'clean' | 'reported'
//   report_count: 3,
//   severity: 'high',          // 'none' | 'low' | 'medium' | 'high' | 'critical'
//   reasons: ['prompt_injection', 'data_exfiltration'],
//   recommendation: 'BLOCK',   // 'CLEAN' | 'MONITOR' | 'CAUTION' | 'BLOCK'
//   latest_report: '2026-03-22T13:49:59Z',
//   reports: [...]
// }

isReported(agent, options?)

Simple boolean check.

import { isReported } from 'kya-abuse-check';

if (await isReported('untrusted-bot')) {
  throw new Error('Agent has abuse reports');
}

checkAbuseMany(agents, options?)

Check multiple agents in parallel.

import { checkAbuseMany } from 'kya-abuse-check';

const results = await checkAbuseMany(['agent-a', 'agent-b', 'agent-c']);
for (const [name, result] of results) {
  console.log(name, result.recommendation);
}

reportAbuse(input, options?)

Report an agent for abusive behaviour. v1.1.0+

import { reportAbuse } from 'kya-abuse-check';

const result = await reportAbuse({
  agent: 'malicious-bot',
  reason: 'data_exfiltration',
  severity: 'high',
  evidence: 'Agent attempted to read /etc/passwd via MCP tool call',
});

if (result.success) {
  console.log('Reported:', result.report.id);
}

Reasons: data_exfiltration, prompt_injection, unauthorized_access, spam, impersonation, malicious_code, other.

Severity: low, medium (default), high, critical.

Rate limited to 10 reports per hour per IP.

Report Reasons

| Reason | Description | |--------|-------------| | data_exfiltration | Agent attempted to steal or leak data | | prompt_injection | Agent injected malicious instructions | | unauthorized_access | Agent accessed resources beyond its scope | | spam | Agent engaged in spam or abuse | | impersonation | Agent impersonated another agent or human | | malicious_code | Agent executed or deployed malicious code | | other | Other abuse not covered above |

Recommendations

| Level | Meaning | Suggested Action | |-------|---------|-----------------| | CLEAN | No reports | Allow | | MONITOR | 1 report | Allow with logging | | CAUTION | 2+ reports or high severity | Allow with restrictions | | BLOCK | 5+ reports or critical severity | Deny access |

How It Works

KYA maintains a community-driven abuse database. Anyone can report an agent. Anyone can check an agent. The more people report, the safer everyone is.

• Free. No API key. No rate limits on checks. No signup.
• Read + Write. Check agents AND report bad ones from the same package.
• Fail-open. If the API is unreachable, agents are assumed clean. Security shouldn't break your app.
• Privacy-respecting. Reports are about agent identifiers, not personal data.

Works With

• mcp-trust-guard — MCP server security middleware with built-in abuse checking
• @agentscore-xyz/mcp-server — MCP server for AI agent verification
• agent-recall — Persistent memory for AI agents

License

MIT