Nvidia open-sources Auto Ontology, a reasoning layer for agents that query enterprise data

Nvidia pushed Auto Ontology to GitHub on 3 November. It is an Apache 2.0 Python library that sits between an agent and the databases it queries. The project's own README frames the problem bluntly. Agents can already fetch documents and run queries, but a revenue number that arrives fast and wrong is worth nothing.

The library generates an ontology from enterprise metadata. It reconciles definitions that disagree across systems. It then exposes the result to "any agent harness" through MCP, a command line interface and an API. According to the README, it complements semantic models a customer already owns rather than replacing them.

"Validate: check generated queries against certified definitions and lineage; return provenance with every answer," the documentation states. The learn loop folds certified corrections back into the ontology. The shared context therefore changes with use.

The stack listed in the repository is ordinary enough to audit. Postgres holds the ontology next to the catalog it describes. The backend runs FastAPI on Python 3.12 or later, with LangChain handling natural language to SQL. The frontend is Next.js 16 with React 19, Prisma and Better Auth. Nvidia NIM powers natural language querying. Deployment is either a Helm chart for Kubernetes or Docker Compose for a local stack. The project is not accepting external contributions. That puts it closer to a reference implementation than a community fork.

That matters for anyone building a fine-tuning or evaluation pipeline. A tuned model that writes SQL is only as good as the definitions it was trained against. The repository points readers to a separate ontology SQL evaluation repository for running benchmarks such as BIRD.

Agents on a leash, and a school for breaking them

The context problem is not Nvidia's alone. Cisco open sourced Project Quarry on 28 October. It is a behaviour detection engine that watches whole agent sessions rather than single tool calls. Amy Chang, head of AI threat and security research at Cisco, told The Register in an exclusive interview that most detection is missing "the entire context under which something may occur." She compared it to card fraud. One odd transaction may pass, but a sequence of them tells a story.

Quarry covers single and multi turn conversations, tool use, Model Context Protocol interactions and multimodal inputs. It ships with plugins for Claude Code, Codex and other MCP compatible agents. The repository includes rule-writing guidance and a starter pack. That is the closest thing to a curriculum the field currently offers.

For a more direct one, the AI Red Team Academy published a free, open source course on breaking AI systems on 3 November. It is a course, not a product. No enrolment figures are public.

Tooling for running several coding agents at once also matured this week. Friday, posted to GitHub on 1 November, drives the coding tools already signed in on a machine. These include Claude Code, Codex, Cursor, Gemini CLI and GitHub Copilot. It moves a task to the next agent when one hits its usage limit. It needs no API keys and no account. Its README says the only data sent back is a few anonymous usage counts that can be turned off.

Smaller projects fill the gaps around the edges. Octomind, an AI harness written in Rust that claims to save tokens, was updated on 3 November. Local Flow, a macOS dictation app that runs Whisper through whisper.cpp with Metal, was updated the same day. Neither ships benchmarks, so treat the token claim as the author's.

The money question nobody open sourced

Whether any of this pays for itself is a separate argument. A study published on 3 November asks directly whether hosting an open source inference engine is profitable. The study poses the question and does not answer it.

The pressure is visible elsewhere in the stack. OpenAI said on 3 November that it is quadrupling bug bounty payouts. Critical vulnerabilities in ChatGPT or Codex are worth up to $30,000, and exceptional findings up to $100,000. Bonuses lift the ceiling to roughly $172,000, according to The Decoder. Anyone who reported a flaw in the previous 30 days gets the difference retroactively.

OpenAI is also rewriting its own vocabulary. The Verge reported on 3 November that the company has begun calling data centers "compute campuses." The term appeared in a roughly 900 word announcement for its Grid Protectors Program, while "data center" turned up twice. Holly Elmore of PauseAI US told The Verge that the name is not what makes the facilities unpopular.

"Rebranding them 'compute campuses' may fool listeners once or twice but it's not the name 'data center' that is unpopular," Elmore said.

None of this settles what an open source fine-tuning stack should look like in production. But the week's releases point in one direction. The model is the cheap part. The definitions, provenance and evaluation around it are where the work, and the scrutiny, now sit.