Argos_

Who it is for

A company of fifty to two hundred people has Microsoft 365, a remote management tool, a backup appliance and no dedicated security role. The work is done by one person who also does other things, spread across spreadsheets, each vendor’s console and email. Argos gives that person one place where the inventory keeps itself up to date, where a new vulnerability is matched against that inventory without anyone asking, and where audit evidence comes from the same data used the rest of the year.

It installs on the company’s own server, runs as a single process against a PostgreSQL database, and is built as modules: fifteen, each switched on separately, so a company with nothing but Microsoft 365 can use the tenant posture measurement without fleet management.

What it does

Inventory. An enterprise CMDB fed by four automatic synchronisations, which remembers the fields a person has corrected so no sync overwrites them.

Scans. Fifteen local tools chained against the CMDB’s own assets, with active scanning. A domain is only scanned if its ownership is verified through a DNS record.

Supply chain. Every dependency of the organisation’s repositories cross-checked against four sources. Higía proposes the fix, tests it in isolation and opens a pull request that a person merges.

Backups. One engine places every backup in a circular week and warns of conflicts; another walks the resource graph and checks the 3-2-1 rule.

Compliance. Statement of applicability and maturity per measure of Spain’s National Security Framework, master plan, risks and threat modelling on the CMDB’s own diagrams.

Talos. The hardening audit came out of here and is published separately: read-only, scores from 0 to 100, and Argos ingests the result.

Plus phishing simulations with training, Microsoft 365 tenant posture, fleet agents, a register of the AI in use under the EU AI Act, an open API with an MCP server and signed licences.

Éaco

It classifies with a language model the emails staff forward to the reporting mailbox and picks the reply from a deterministic matrix. Its input is text written by an attacker and its output can end up in an email under the security team’s letterhead: indirect prompt injection with an output channel. The defence is not a better prompt, it is the boundary. The model can only return one type from a closed list of twelve, a confidence and a reason, and the reason never travels into an email. The worst an attacker achieves is having their message misclassified.

Next to it, Proteo: an adversarial bench against the in-house agents, with prompt injection, jailbreak, system prompt extraction and tool abuse, measuring the success rate of each attack.

The guards

There are 172 test files whose only job is to turn red when someone breaks a convention: no route resolves the session by hand, every mutation leaves a trace in the audit log, every scheduled job declares its time zone.

In the adversarial round closing one release I found four that were green without measuring anything. One detected itself through a comment. Another carried a regular expression missing its backslashes. A third had an assertion that could not fail. The fourth declared a behaviour it did not have. Hence the project rule: a guard nobody has seen turn red is a hypothesis.

With agents

I build a platform this size by directing AI agents, with the rules written as files: the architecture, the boundaries of each module and what each one cannot touch. The design, the judgement and the verification are mine; the volume of code is moved by the agents. 5,093 commits between April and October 2026, 281 data models, 681 API routes and 1,763 test files.

Status

Advanced beta: the engineering done and the operations still to do. No production deployment, and Éaco in shadow mode awaiting a product decision. The code is not public.