Talos
An AI agent with a permission kernel between model and shell
Last verified:
What is Talos?
Talos is an AI agent with a kernel-based security architecture that executes tasks through 23 declared tools (15 read, 5 write, 3 exec) with mandatory audit logging and sandboxing. It enforces a security pipeline (path floor → hardline → dangerous → effect) to prevent unauthorized access, unwanted commands, and file mutations, and supports multiple interaction channels including terminal, Telegram, IMAP mail, and WhatsApp.
Talos pricing
Pricing model: Freemium
Talos pros
- Security-first kernel design with manifest-based tool declarations—undeclared tools are refused on sight
- Multi-channel operation (terminal, Telegram, mail, WhatsApp) with fetch-only inbound communication
- Deterministic control mechanisms (/stop, /undo, /autonomy levels, /log) that operate without language model intermediaries
- Sandboxed code delegation and browser automation for bounded coding tasks with kernel-derived workspace isolation
Talos cons
- Requires platform-provided sandboxing (bubblewrap on Linux, sandbox-exec on macOS); refuses to run without it
- Single-operator, single-machine design—not designed for multi-tenant environments
- Does not defend against intentionally malicious models; only against mistakes and injection attacks through tool output
Frequently asked questions about Talos
How does Talos prevent dangerous commands from running?
Commands pass through a 'hardline' check against protected paths. Clean commands run sandboxed and logged. Risky operations (like writing to secrets or system files) ask for operator confirmation before execution.
What communication channels does Talos support?
Terminal, Telegram, mail over IMAP, and WhatsApp through an operator-controlled broker over SSH. All inbound communication fetches data—nothing opens a listening socket.
Can Talos defend against a malicious AI model?
No. Talos defends against mistaken commands and injection attacks through tool output, not against intentionally malicious model behavior. It's designed for safety within a trusted model context.