1. Security model
ArgSeal separates model intelligence from execution authority. The host captures what the user approved, issues opaque authority handles for protected values, validates each protected field at the tool boundary, and resolves sensitive values only after deterministic admission succeeds.
The model may propose an action or provide advisory analysis. It cannot mint authority, alter host registries, override deterministic policy, or authorize protected execution through a provenance claim.
Current enforcement includes tool, field, run, issuer, expiration, revocation, and single-use grant checks. Runtime uncertainty and internal security-path failures are designed to block rather than permit execution.
2. Current tested support
- OpenAI Agents SDK custom function tools wrapped with ArgSeal
protectTool. - Configured protected fields represented by host-issued opaque references.
- Host-only registries, guarded protected-value resolution, and call-bound admission handoff.
- Redacted security spans and bounded execution records.
- Current tested dependency:
@openai/agents 0.13.5. - Synthetic no-network email sink and fixed evaluation scenarios.
3. Known limitations
The current MVP does not claim protection for:
- Hosted tools, hosted MCP tools, native handoffs, or
agent.asTool(). - Web search, file search, computer use, shell, or patch tools.
- Production email, payment, storage, file-sharing, or workflow connectors.
- Arbitrary agent frameworks without a tested ArgSeal adapter.
- Actions or effects that occur outside an instrumented ArgSeal path.
- Security defects in surrounding applications, deployment configuration, dependencies, identity systems, networks, browsers, or provider infrastructure.
- Absolute prevention of every attack or implementation defect.
The public Lab is a demonstration, not a penetration-testing target, managed security service, certification, or production deployment.
4. Reporting a vulnerability
Send security reports to [email protected] with the subject line Security report.
Please include:
- A clear description of the issue and its security impact.
- The affected URL, component, version, commit, or configuration.
- Reproduction steps using synthetic or redacted data.
- Observed and expected behavior.
- Any relevant logs, screenshots, or proof-of-concept material that does not expose secrets or third-party data.
- Your preferred contact details and disclosure timeline.
Do not send API keys, access tokens, private keys, real customer data, medical data, financial data, employee data, or other sensitive information. Redact secrets before submission.
5. Good-faith research guidelines
We support responsible, good-faith research that is limited to systems you are authorized to test and that avoids harm. When evaluating the public ArgSeal services:
- Use the fixed synthetic scenarios and normal user-facing interfaces.
- Do not access, modify, delete, or retain another person's data.
- Do not degrade availability, generate excessive traffic, or bypass rate limits.
- Do not use social engineering, physical attacks, denial of service, malware, credential stuffing, or attacks against employees, providers, or unrelated infrastructure.
- Stop testing and report the issue if you encounter secrets, non-public data, or a path that could cause real external effects.
- Allow reasonable time for investigation and remediation before public disclosure.
These guidelines do not authorize testing of Cloudflare, OpenAI, GitHub, or any other third party. Follow each provider's own security policy and obtain permission where required.
6. What to expect
We will make a reasonable effort to acknowledge a complete security report, assess severity, request additional information when needed, and communicate material remediation progress. Response time may vary based on project stage, complexity, and available resources.
We may decline reports that are out of scope, purely theoretical without meaningful impact, duplicate known issues, dependent on unsupported integrations, or based on automated scanner output without validation.
7. Bug bounty status
ArgSeal does not currently operate a paid bug bounty program. Submission of a report does not create an entitlement to payment, credit, employment, or a commercial relationship. We may choose to acknowledge helpful reports at our discretion.
8. Contact
Security reports: [email protected]
Legal questions: [email protected]