Inspiration
Coding agents now delete files, read secrets and run commands. Humans cannot watch every step, and clicking "approve" on everything defeats the purpose. Banks solved the same problem decades ago without watching every cashier: nobody both makes and approves an entry, every teller has a limit, everything goes into an append-only ledger, and the books are reconciled at the end of the day. Talos ports those controls to AI agents.
What it does
Talos sits between a coding agent and its tools. Every action passes through a deterministic risk engine: LOW actions run; MEDIUM actions are reviewed by an independent Checker model; HIGH actions wait for a human; hard blocks never run. Everything is written to a tamper-evident, hash-chained ledger. At the end of the day Talos reconciles what you asked, what the agent planned and what actually happened, posts mismatches to suspense, can roll back any write, and issues a sealed day-close certificate that anyone holding the ledger can verify.
Features: maker-checker with a context-isolated, fail-closed Checker; teller risk limits that freeze a misbehaving agent; hash-chained ledger; three-way reconciliation with suspense; rollback; secret taint tracking; bank-style authority levels (Trainee, Standard, Senior); a "threats stopped" view; a session replay slider; and the sealed certificate.
How we built it
Python and Streamlit. Talos is a library plus a dashboard: the Gate wraps every tool call, the RiskEngine classifies it, the Teller meters it, the Checker reviews it, and the Ledger records it. 58 automated tests cover tamper and truncation detection, taint tracking, fail-closed behaviour, rollback conflicts, reconciliation drift and certificate forgery. The agent's executor has no shell, a scrubbed environment and no network, as defence in depth.
Nebius Token Factory and NVIDIA Nemotron. The design uses two different models through Token Factory's OpenAI-compatible API: Nemotron 3 Super as the Maker (runs the coding-agent loop) and a different, larger Nemotron model (Ultra) as the Checker. The Checker sees only the structured action and the original task, never file contents or the Maker's reasoning, so a prompt injection hidden in a README cannot reach it. It fails closed: an API error or an unparseable answer means reject. Talos refuses to start if Maker and Checker are the same model (four-eyes rule), and the large model is called only for risky actions.
Honest status of live mode. The Token Factory integration is implemented and unit-tested against a fake client, but I could not run it live: Bangladesh is not in the Token Factory billing country list, so I could not finish onboarding or redeem the hackathon promo code. Nebius Support pointed me to Discord, where I asked for help and had not received an answer by the time of submission. Because of this, the demo video and the hosted demo use the offline simulation (a scripted agent and a heuristic Checker), and are labelled as such. Everything except the model calls (risk engine, teller, ledger, reconciliation, rollback, certificate) is the real code path.
Challenges we ran into
- Making the Checker useful without letting an injected instruction reach it. Solved by passing it only structured facts, never file contents.
- A hash chain alone cannot show that the newest entries were deleted, so the ledger exposes an anchor (head() / verify(anchor)) and the README states this limit.
- Blocked attempts had to cost budget too, otherwise a stubborn agent could probe forever for free.
- Not being able to reach the live model endpoint from Bangladesh, which pushed me to make the whole control layer testable and demonstrable offline.
Accomplishments that we're proud of
In the demo, an agent hijacked by a poisoned README tries to steal a .env file: the credential read is held for a human, the outbound POST carrying secret material is hard-blocked, and the curl fallback is blocked as a network tool. The whole session can then be rolled back and sealed in a certificate that fails verification if the ledger, the sealed data or even the visible text is edited.
What we learned
Safety for agents does not need to mean watching every step. Borrowing mature controls from banking (separation of duties, limits, append-only records, end-of-day reconciliation) gives a simple, explainable structure. I also learned that a control layer should be testable without any model: deterministic rules do most of the work, and the model Checker adds an independent second opinion on top.
What's next for Talos: dual control for AI agents
Run the full scenario live on Nemotron models as soon as Token Factory access is available, then plug Talos into real agents (Claude Code hooks, Cursor, Codex), add per-team policy files and multi-session dashboards.
Built With
- nebius-token-factory
- nvidia-nemotron
- streamlit
- ython