Hackathons

Hackathons

Public builds made for hackathons. Each entry is published with what it shows, what was measured on its own corpus, and what it does not claim.

01

Entry

BlastGuard AI: AI that can refuse, with proof

Nebius x NVIDIA hackathon on Devpost, submitted 30 October 2026

What it shows

Because a secure patch can still break production. "Findings: 0" is not the same as "still works".

An agent loop finds misconfigurations, asks NVIDIA Nemotron for a YAML patch, re-verifies from scratch, then an Intent Preservation Gate returns SAFE_TO_PROPOSE or REVIEW_REQUIRED. The model only proposes; checks that do not depend on the model decide.

Still from the BlastGuard AI demo: runtime proof on a kind cluster
Demo video — opens on YouTube in a new tab.
02

Pipeline

Find → patch → re-verify → gate.

Intent Preservation Gate after deterministic rules and external validators.

BlastGuard remediation pipeline Find misconfigurations, ask Nemotron for a patch, re-verify up to three attempts, then an Intent Preservation Gate branches to SAFE_TO_PROPOSE or REVIEW_REQUIRED. Find rules + validators Patch Nemotron propose Re-verify up to 3 attempts Gate intent check SAFE_TO_PROPOSE REVIEW_REQUIRED BlastGuard remediation pipeline Find misconfigurations, ask Nemotron for a patch, re-verify up to three attempts, then an Intent Preservation Gate branches to SAFE_TO_PROPOSE or REVIEW_REQUIRED. Find rules + validators Patch Nemotron propose Re-verify up to 3 attempts Gate intent check SAFE_TO_PROPOSE REVIEW_REQUIRED
03

Measured

On its own small corpus.

Small, hand-written by one author, static oracles — published as is.

On its own small corpus · 16 cases, 48 real runs

A security-only tool would accept all 35. BlastGuard sent those 14 to review, and sent 0 function-preserving patches to review.

0/14 of 14 function-breaking, security-passing patches accepted as safe
6/6 already-clean manifests needed no change
6/6 invalid inputs rejected before any model call
3/3 prompt injections inside a manifest ignored
04

Runtime proof

One synthetic case in kind.

Security-passing patch that stopped collecting; gate refused.

  1. Original COLLECTED
  2. Nemotron patch NOTHING_TO_COLLECT passes every security check
  3. BlastGuard gate REVIEW_REQUIRED

Kind cluster in GitHub Actions: synthetic collector on a node hostPath. Attempt 1 failed (kept hostPath); attempt 2 passed security; gate returned REVIEW_REQUIRED.

05

Limits

What it does not claim.

Blind challenge on this corpus: 6 unseen cases, 18 real runs — frozen oracles found 0 false SAFE and the gate sent 13 of 18 to review.

A later cluster audit confirmed 2 failure modes across 6 runs; 3 could not be tested and 4 may be false alarms, so the oracles were too lax and the gate is strict but its reasons are generic.

  • VERIFIED means "passed these checks", not "works in your cluster".
  • The app never touches a cluster.
  • Corpora are small and hand-written by one author; oracles are static and conceptually overlap the gate.
  • The runtime proof is one synthetic case.
06

Links

Built with: Python, Streamlit, PyYAML, openai SDK, NVIDIA Nemotron (nemotron-3-super-120b-a12b), Nebius Token Factory, kubeconform, kube-linter, GitHub Actions, kind.

Provenance: the original BlastGuard (Go CVE blast-radius engine) existed before the submission period and is not integrated; everything in blastguard-ai was built during the submission period.

Build details on this setup
Parsing
Fail-closed YAML (safe loader, size and document caps; manifests never executed)
Rules
13 deterministic rules BG001–BG013
Validators
kubeconform schema validation and kube-linter
Sandbox
Patch applied in a sandbox; identity guard blocks deletes/swaps of the workload
Feedback
Verifier feedback returns to the model; parse, API, budget or validator problems fail closed
Demo caps
Per-session 10 and process-wide 1500 run limits

How these pages are written

Every number on this page comes from the project's own published evidence.

Available for meaningful infrastructure conversations

Build systems that stay calm under pressure.

Available for conversations

Talk infrastructure

Book a working session

Meet ↗

Evidence first

For deep dives

Explore the systems

All cases ↗