Placed 2nd in the Google DeepMind track at the RAISE Summit hackathon, Paris.

Submission

The RAISE Summit hackathon ran the first weekend of July 2026, produced by Cerebral Valley around the Paris summit and billed as the largest AI hackathon in the world, with winners announced per sponsor track. We entered the Google DeepMind track, whose brief asked for applications running Gemma locally for offline, privacy-first inference, and we competed remotely: the track is judged entirely on a demo video, the public repository, and the written description, with the demo weighted at half the score.

We submitted Airlock, an on-device firewall for AI agents. It screens the untrusted content an agent is about to read and quarantines prompt injection and data exfiltration before either reaches the model's context, with every verdict computed by Gemma through Ollama on the operator's own machine. That property is what the track existed to surface: a screening layer that uploads what it screens would be carrying the very secrets and attacks it exists to contain. The rules required new work only, built in the open, and the repository's public commit history is the compliance record.

Result

Results were announced two days later. Airlock placed 2nd in the track. The judging was remote by construction, which means the placement rests on exactly three artifacts: a sixty-second demo video, the public repository, and the submission's written description. Nothing about the result depended on presence in the room.

The result reads as operational signal in the same sense the San Francisco showcase did. The brief came from Google DeepMind, the models were Gemma, and the published weighting put half the score on the demonstrated behavior of the system. What placed was the working mechanism, not a pitch.

What changed

Airlock joined the lab's releases the week the result landed, with its own page and a maintained public repository. The research thread it opens is the input side of agent reliability. The lab's existing work concerns what an agent remembers and how it verifies what it ships; Airlock adds the discipline of what an agent may read. The three constraints compound, because an agent trusted with autonomy needs all of them at once.

What did not change

The lab's positioning did not change, as it did not change in February. Hackathons are not a strategy; they are a place where a constraint the lab already carries can be built against a deadline and judged by people with no stake in being kind. The cadence of research, deployments, releases, and writing continues in the same order, and the next piece of work is judged on its own merits.

Standing reference

This note is the canonical reference for the result. Airlock is documented on its own page and ships as X-Arc-ai/airlock on GitHub under Apache 2.0. The placement sits alongside the February recognition as the second time this year the lab's work has placed inside a frontier lab's own brief.