owthorize

Prompt safeguards are theater. The model isn't the boundary — the layer between it and your systems is.

owthorize is a synchronous JS/TS gate that catches destructive AI-agent tool calls before they execute, using ASTs and parsed shapes — not regex on strings. It sits at the tool layer of your agent (OpenAI, Anthropic, LangChain, Vercel AI SDK) and gates every call through a small rule engine.

// the model emits this:
db.query("DROP TABLE users")
//              │
//              ▼  guard.tool() intercepts, parses, denies
//              │
throw new GuardDenied("DDL not allowed: drop")

Modern AI agents call tools. Tools have side effects. Three things go wrong:

1. Prompt injection. A user (or a document the agent reads) coerces the model into emitting calls the developer never anticipated.
2. Hallucinated arguments. The model forms a syntactically-valid call with semantically-wrong inputs — DELETE FROM users instead of DELETE FROM users WHERE id = ?.
3. Reasoning errors. The model tries to "be helpful" by issuing destructive cleanup, fetches an internal IP it shouldn't reach, or writes outside the workspace.

Prompt-level safeguards ("you are a helpful assistant, do not drop tables") fail under all three. The right boundary is the layer between the model and your database/HTTP/disk — the tooling layer. That's where owthorize lives.

• SQL DDL — DROP, TRUNCATE, ALTER, CREATE, RENAME (Postgres / MySQL / SQLite, AST-level)
• Unbounded mutations — UPDATE / DELETE with no WHERE clause
• SSRF targets — RFC1918, link-local, loopback, AWS metadata 169.254.169.254, IPv4-mapped IPv6 ([::ffff:127.0.0.1]), *.internal wildcards
• Shell metacharacters and dangerous commands — basename match (/usr/bin/rm → rm), pipe / redirect / $() / backtick detection
• Path traversal — resolved-path containment, prefix-collision-safe (root /safe doesn't match /safe-evil)
• Project-specific policy — typed custom rules per adapter (rules.sql.custom, rules.http.custom, etc.)

JS/TS developers shipping AI agents that touch:

• a database (Postgres, MySQL, SQLite — via mysql2, pg, sqlite3, etc.)
• outbound HTTP (fetch, axios, undici)
• the filesystem (fs/promises)
• shell exec (child_process)

If your agent is read-only against trusted data, you probably don't need this. If it can write — or if it can fetch arbitrary URLs — owthorize closes the most-likely failure modes before they ship.

Both ESM and CJS are supported. Node ≥ 18.

import { Guard, rules, GuardDenied } from "owthorize"

const guard = new Guard({
  rules: [
    rules.sql.denyDDL(),
    rules.sql.denyMutationWithoutWhere(),
    rules.http.denyHosts(rules.http.SSRF_DEFAULTS),
  ],
})

const safeQuery = guard.tool("db.query", {
  adapter: "sql.postgres",
  handler: async ({ query }: { query: string }) => db.query(query),
})

try {
  await safeQuery({ query: "DROP TABLE users" })
} catch (err) {
  if (err instanceof GuardDenied) {
    console.log(err.matched, "→", err.reason)
    // → sql.denyDDL → DDL not allowed: drop
  }
}

await safeQuery({ query: "SELECT id FROM users WHERE id = $1" })
// runs the handler

Test rules without side effects:

guard.simulate("db.query", { query: "DROP TABLE users" })
// → { decision: "deny", reason: "DDL not allowed: drop", matched: "sql.denyDDL", irreversible: true }

For the full guide (custom rules, framework shims, audit log, failure modes, the irreversible flag): see USAGE.md.

Adapters parse the raw payload into a typed shape. Rules see ASTs, not strings.

protectTools(guard, tools, perTool?) wraps an entire tool registry in one call:

import { protectTools } from "owthorize/openai"

const safeTools = protectTools(guard, openaiTools, {
  db_query: { adapter: "sql.postgres", redact: ["params.password"] },
})

Available at:

• owthorize/openai — Array<{ type: "function", function, handler? }>
• owthorize/anthropic — Array<{ name, description, input_schema, handler? }>
• owthorize/langchain — Array<{ name, description, schema, func | _call }>
• owthorize/vercel-ai — Record<string, { description, parameters, execute? }>

All four shims preserve framework-specific fields (strict, experimental_*, etc.) and pass schema-only tools (no handler / no execute) through untouched.

Every check writes a structured record:

{
  "ts": "2026-05-01T12:00:00Z",
  "tool": "db.query",
  "adapter": "sql.postgres",
  "parsed": { "kind": "ddl", "ddlOp": "drop", "tables": ["users"], "hasWhere": false },
  "payload_hash": "sha256:898f9f1f624ca10a59e92c9483c39a9991f5664d19135b4b442fd92e1c92cb1d",
  "decision": "deny",
  "matched_rule": "sql.denyDDL",
  "matched_rule_kind": "builtin",
  "reason": "DDL not allowed: drop",
  "irreversible": true,
  "simulated": false
}

• Pluggable sink (audit: { sink: myLogger })
• silentSink exported for tests / dev
• Field-level redaction (redact: ["params.password"]) — sensitive values are stripped before hashing, so the hash is stable across rotated secrets
• Audit failures fall back to a secondary sink and never break tool execution

Built-in rules tag denies that block actions you couldn't easily roll back (DDL, unbounded mutations, destructive shell, writes outside fs roots). Custom rules opt in via deny(reason, matched, { irreversible: true }). Consumers route on it — auto-deny most things, escalate the irreversible ones to a human:

const decision = guard.simulate("db.query", payload)

if (decision.decision === "allow") return safeQuery(payload)
if (decision.irreversible) return mySlackBot.requestApproval(decision)
return res.status(403).json({ matched: decision.matched })

The SDK never blocks waiting for approval — it returns synchronously and lets your code decide whether to gate, route, or deny.

Catches: prompt-injected tool calls, hallucinated arguments, agent reasoning errors, classes of unsafe shapes (DDL, unbounded mutations, SSRF, shell metacharacters, path traversal).

Does not catch: a malicious agent runtime that bypasses the SDK, vulnerabilities inside your tool implementations, side effects that occur before the tool boundary. The trust boundary is the wrap; what you don't wrap, you don't gate. For defense against a hostile runtime, you need a process boundary (proxy, sidecar, container egress rules) — that's a different product.

1. Parse, don't match. Rules see ASTs and parsed URLs, not strings. Regex on SQL is a defect generator.
2. Synchronous v1. Allow or deny, return immediately. No webhooks, no async resume. The irreversible flag lets consumers route to their own approval flow without coupling the SDK to one.
3. Default deny on uncertainty. Unknown tool, parser failure, rule exception — all deny by default. Each is configurable.
4. Wrap once at the tool layer. guard.tool(name, handler) is the API. Developers don't branch on decision.status === "allow" in their own code.
5. Testable from day one. Built-in rules ship with unit tests. simulate() is a first-class API.

owthorize is at v0.4.1 — the public API is stable but the version stays sub-1.0 until field feedback from outside the original author lands. The validation log lives in FIELD-TESTING.md and the running paper-cut report in field-report.md. Both have been re-tested end-to-end across all four adapters, framework shims (OpenAI + Vercel AI against gpt-4o-mini), and the irreversible flag on a real Express + Drizzle + MySQL backend.

On the path to v1.0:

• API stabilization based on real-world feedback from external users
• LangChain + Anthropic shim end-to-end runs (built and unit-tested; not yet field-validated)
• Approval-flow recipes (Slack / queue) as documented patterns, not SDK code

Explicitly not on the roadmap (these would deform the synchronous v1):

• Built-in human-approval UI / state machine
• Hosted policy server, multi-tenant control plane
• Anything async beyond the handler itself
• Row-count / blast-radius estimation that requires running the query

• USAGE.md — full user guide (custom rules, framework wiring, audit, failure modes)
• MVP.md — original v1 spec
• FIELD-TESTING.md — validation status of every public surface
• field-report.md — running log of what's been dogfooded against real traffic

Issues and PRs welcome. Bug reports with a minimal reproduction (the call site that surprised you, plus the audit-log line if you have it) are gold.

git clone https://github.com/Spyyy004/owthorize.git
cd owthorize
npm install
npm run typecheck
npm run lint
npm test
npm run build

The example scripts (npm run example, npm run example:openai, npm run example:vercel-ai) require an OPENAI_API_KEY for the framework-shim ones. The plain quickstart runs without any API key.

MIT — see LICENSE.