AaaS handles private provider sessions, local source snapshots, Runner credentials, and remote execution. Please report security issues privately and avoid putting real user data at risk while testing.
| Version | Support |
|---|---|
main |
Active development |
0.2.x |
Public preview fixes |
< 0.2 |
Not supported |
The hosted public control plane is a functional preview and has no uptime, support, or vulnerability-reward SLA.
Use this repository's private GitHub security-advisory flow. Do not open a public issue.
Include:
- the affected version or commit;
- the affected surface and execution mode;
- a minimal reproduction using synthetic data;
- the security impact and realistic attack prerequisites;
- whether source sessions, capsules, credentials, or other users are exposed;
- a suggested mitigation, if known.
Do not include a live provider key, Runner token, private session file, transcript, local home path, or data belonging to another person.
Maintainers make a best-effort attempt to:
- acknowledge a complete report within 3 business days;
- provide an initial severity assessment within 7 business days;
- coordinate remediation and disclosure based on impact and exploitability.
Complex runtime or provider issues may take longer. Please allow a reasonable remediation window before public disclosure.
- installer role escalation or unwanted Runner installation;
- source-session or transcript disclosure;
- cross-conversation access or writes;
- Runner authentication, lease, or job-claim bypass;
- encrypted capsule confidentiality or integrity failures;
- control-plane authorization or data-isolation failures;
- secret exposure through logs, errors, manifests, or package contents;
- sandbox or runtime isolation bypass specific to AaaS.
- vulnerabilities in Codex, Claude Code, Cloudflare, Sites, Node.js, or another third-party product that do not arise from AaaS;
- availability of a publisher-operated local Runner;
- missing accounts, quotas, billing, or anti-abuse features already disclosed as preview limitations;
- automated scan output without manual validation and concrete impact;
- social engineering, physical access, denial of service, or testing that accesses another user's data.
- Use synthetic sessions, Agents, Conversations, and credentials.
- Do not degrade the public service or run indiscriminate scanners.
- Stop if testing could expose or modify data you do not own.
- Prefer a local control plane and isolated Runner for proof of concept.
- Remove sensitive values and identifying paths from screenshots and logs.
The control plane stores public metadata, Agent and Conversation state, messages, job leases, Runner token hashes, and encrypted cloud capsules. Local source paths, raw provider session files, and provider credentials remain on the Runner.
Cloud capsule encryption protects stored capsule content and callers without the Runner secret. The preview currently trusts the control-plane runtime and the paired Runner host. Write-enabled tools require additional per-conversation workspace, identity, secret, network, and approval isolation.
See Architecture for the complete model.