Current public release candidate: v0.11.0rc2
CATL is open source under the Apache License 2.0 and is designed as a defensive trust layer for AI-agent and MCP tool execution. It is developer-tested security software, not an independent security certification.
The protected runtime currently supported by the release candidate is Linux. The intended deployment separates broker, agent, human operator and tool-backend authority across distinct identities and authenticated Unix-socket channels.
Use the protected broker path for consequential testing:
catl brokercatl backend-servicecatl shimcatl overlay --operator-socket ...
The deliberately unsafe demonstration paths, including catl run --allow-unsandboxed and overlay --unsafe-direct-state, are for disposable evaluation only.
- Default deny: unknown or unauthorised tool actions do not silently pass.
- No silent trust reset: anchored state must not be automatically reinitialised after integrity failure.
- Separated authority: broker policy/signing state must not be writable by the untrusted agent.
- Exact approvals: operator approvals are action-bound and replay-resistant.
- No blind retry of uncertain effects: externally ambiguous operations are quarantined for reconciliation.
- Host controls still matter: CATL cannot protect a capability the agent can reach through an unmediated shell, plugin, credential, network route or alternate tool registry.
- No mandatory cloud control plane: CATL does not require a listening public TCP API, automatic updater or external account for local trust decisions.
Please do not open a public issue containing exploit details, secrets, customer data or a working bypass.
Preferred route: use GitHub's private security-advisory reporting flow for this repository:
https://github.com/BryanFiFife/CATL/security/advisories/new
If private advisory reporting is unavailable, contact Caledapt through its established business contact channel and clearly mark the message CATL SECURITY.
A useful report includes:
- affected CATL version/commit;
- platform and deployment mode;
- minimal reproducer;
- expected vs actual trust-boundary behaviour;
- impact and realistic attacker prerequisites;
- any proposed mitigation.
For v0.11.0rc2, public evidence includes:
- 31/31 installer regression tests;
- 1,000/1,000 hostile-manifest cases rejected;
- disposable Linux install → verify → reinstall smoke testing;
- static
systemd-analyze verifychecks on rendered broker/backend units; - offline systemd security analysis scoring 1.3 / OK for both reviewed service profiles;
- the prior core baseline recording 278 passing Linux regressions;
- a four-identity signed rootless E2E covering forgery denial, tool execution, approval/replay control, audit verification, channel isolation and broker-key denial.
See:
- rc2 release notes
- installer acceptance
- threat model
- rootless Linux deployment
- v0.11 red-team assessment
- release attestations
CATL does not claim:
- protected Windows AppContainer or macOS sandbox operation end-to-end;
- completed hardware-backed FIDO2/WebAuthn presence assurance;
- independent penetration-test certification;
- an independently administered off-host witness;
- proof that an external provider actually settled an email/API/payment effect merely because a tool returned success;
- protection against a compromised host administrator/kernel or stolen operator authority.
Security changes should include a reproducer and regression test wherever practical.