Blog

Vigil 1.1: Emerging as the most popular AND most functional AI SOC

|

The takeoff in Vigil adoption has made it arguably the most widely deployed AI SOC. As I write this, the cumulative clone count is approaching 200,000, much of it in the last several weeks. The figures are public at github.com/Vigil-SOC/metrics.

All this attention, feedback and community support, together with our investment in top software engineering and the Vigil factory, now shows up as accelerated product development.

We scoped a two-week sprint for 1.1. We did more than planned and are releasing half a week early.

As I have said for a while now: thump, thump, the upstream cometh. Once a community starts collaborating, especially with a state-of-the-art and freely available software factory, traditional proprietary vendors have a hard time keeping up. We are reaching that point with Vigil.

Highlights include:

  • Threat intelligence to threat hunts: a community member contributed deeper integration with threat intelligence platforms (TIPs), and operators already use it in ongoing, automated threat hunts.
  • Headless Vigil: we are committed to this increasingly common use case, in which teams use Vigil as the central point to capture learning and enforce controls and rarely open the graphical interface.
  • Federated detection: many DeepTempo friends know us for reinventing threat detection with LogLM, and we also understand that existing rules-based detections and queries are useful, and that an intelligent agentic SOC can improve them.
  • Build your own agents: easier than ever. Security engineers write their own SOC agents as a simple manifest, no code required, and Vigil runs them with the same ledger, approvals and tracing as built-in agents.
  • Safety for autonomy: improvements to the autonomy ceiling, which keeps outside actions off until a person turns them on; custom roles; and the soft launch of Medic, a watcher that keeps an eye on Vigil itself. Keep an eye on Medic; we are investing heavily in this pattern and seeing promising results.

Architecture

Last week I pointed out that an enterprise can outsource a great deal, but, as Satya Nadella and others have argued, it must not outsource what the firm learns. That learning is the firm's core IP; without it the firm loses its differentiation and hence its ability to compete.

Many firms now see cyber as a source of differentiation as well. For firms that agree and want to control how their cyber systems learn, Vigil is becoming central.

In that post I coined the term System of Learning. We built 1.1 around the role Vigil increasingly plays as the System of Learning for cyber.

Vigil holds the agent harness, the ledger and memory of what past runs saw and concluded, the learned confidence score, and the operator's declared intent, kept as Markdown under change control. Vigil works with systems already in place: one or more threat intelligence platforms, one or more SIEMs, and the systems of control that can isolate a host or block a connection.

LogLM, our foundation model for security telemetry, feeds both a SIEM and Vigil. LogLM is optional; it adds the behavioral anomaly detection that a growing number of regulators now expect enterprises to use against today's attacks. I wrote recently about what regulators say about AI-powered attacks and behavioral anomaly detection.

Improved use cases

Threat intelligence to threat hunts

Before 1.1, Vigil typically ingested threat intelligence by reading STIX over TAXII from a single source and using simple indicators (addresses, domains, URLs and hashes), or an analyst uploaded a PDF by hand.

Vigil 1.1 adds an OpenCTI integration, a pattern that applies to any threat intelligence platform. Vigil now ingests a continuous flow of full reports and relies on the TIP for filtering, deduplication and related curation.

An agent first checks whether existing detections and prior hunts already cover a report. If they do not, a hunt agent turns the report into hypotheses, queries the telemetry in place, and writes its findings to the ledger. The next hunt starts from that record. Here the learning, your learning, is plain to see.

If you also run DeepTempo's LogLM, it supplies the hunt with concerning sequences and powers similarity search, which makes your hunts more resilient to shifts in attacker behavior.

Federated detection

Large operators run several SIEMs; one provider we met this month supports seven or eight. Their rules live in those systems, and their detection engineers do not want another place to keep them.

In 1.1 a detection engineer manages the rules an environment runs across Splunk, Elastic and Wazuh (insert your SIEM here) from Vigil, while each rule stays in the system that executes it. An engineer or a tuning agent can lint a candidate rule, replay it against recorded telemetry, and promote it through defined stages, with each change written to the trail.

With LogLM in place, Vigil compares detection layers: a concerning sequence that no rule caught marks a gap worth an engineer's attention. LogLM detects with 99 percent zero-shot accuracy and a false positive rate of 1 percent or less.

Agents you build and run

General coding harnesses lack what a SOC requires of an agent: evidence, authority and memory. In 1.1 an operator writes an agent as a declarative manifest that states its prompt, the tools it needs, its skills, its budgets, and the guard on its conclusions. The manifest carries no code. The operator saves it through the API or the console's Agent Builder, and the kernel runs it under the same ledger, approvals, replay and tracing as the built-in agents.

With 1.1, security analysts can become agent builders without mastering the scaffolding required to run agents reliably in cyber security.

Platform

Headless operation

Many of the engineers adopting Vigil work from a terminal, and several partners deploy on premises where no one watches a screen. In 1.1 the console becomes one client among several.

We have frozen more of the contract: reads of the audit trail and of the autonomy ceiling join the versioned /api/v1 surface and the MCP tools. Vigil now writes approval requests, parked and resumed runs, and rule stage changes to the trail, and pushes them to a webhook as signed CloudEvents. Operators who set an OpenTelemetry endpoint receive traces against a documented span schema. The API documentation also ships inside the image, so it remains available in air-gapped deployments.

Controls

Autonomy is trustworthy only when it is transparent and bounded by strong controls.

Autonomy ceiling. Vigil classifies what an agent may do into five effects: read, write Vigil's own records, propose, act reversibly outside Vigil, and act irreversibly outside Vigil. A new install permits the first three. Only a person holding the autonomy permission can raise the ceiling, and Vigil records that act. The operator can narrow a grant to named tools, a confidence floor or a severity floor. Confidence alone does not release an outside action. One stop control halts unattended runs, and only a person can resume them.

Roles. Administrators build custom roles from one permission registry, starting from presets for a SOC lead, a detection engineer and an auditor. The same permission check governs the console, the API, chat tools, MCP clients and tokens, and it governs reads as well as writes.

Medic. A deterministic watcher with no LLM in its path checks that data arrives from each source, that the model path responds, and that accepted work completes. It records its decisions in an append-only, hash-chained log, and it loads only signed content packs.

LogLM

Each day brings more evidence that attackers use open-weight models to raise both the sophistication and the volume of attacks. A defender who answers by sending each event to a reasoning model will run out of budget before the attacker runs out of attempts. A defender who relies only on signatures drives while looking in the rearview mirror; that approach cannot keep pace with the velocity of novel attacks.

The state-of-the-art architecture pairs an inexpensive, always-on behavioral anomaly detection model, plus rules and signatures, with reasoning where it counts. LogLM is that model. It understands behavior across protocols and telemetry types without a baselining period, and Vigil's agents investigate what it surfaces. Each makes the other more useful, and neither requires telemetry to leave the operator's environment.

This is your community

Vigil is 100 percent open source under Apache 2.0.

Clone it, run it against your own models, and tell us what you build. Community office hours and the documentation are at vigilsoc.org. Please drop into our Discord as well, and let us know if you would like to look at Vigil Assured, the locked-down version we run for customers.

Read more about Assured at deeptempo.ai/vigil-assured.

See the threats your tools can’t.

LogLM works with your existing stack to uncover the evolving threats traditional systems overlook. Vigil, the open source AI SOC, turns those findings into action. No added complexity, nothing replaced.