# Daniel Aguirre

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [DAN AGU](https://cerebralvalley.ai/u/ISSASODA)
- **GitHub:** https://github.com/Ataraxiainc2/ISSAAS
- **Demo video:** https://www.youtube.com/watch?v=dii033mtx-0&list=PLQ1fecUx58p4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/997

ISSA AS (Issue Scouting, Solving & Attestation) is an autonomous agent that finds publicly reported security bugs in open-source Python projects, fixes them, and proves every fix, or abstains. Maintainers are flooded with unverified AI patches; ISSA AS only counts a fix that a deterministic check confirms.

The loop: Muse Spark Open (open weights, on Baseten) reads ISSA AS’s own past outcomes and proposes where to search GitHub next; code validates each proposed search (no repo/user/org targeting) and runs it with a read-only token. Spark Open explains every candidate in plain English, Spanish and Portuguese (many issues are written in Chinese), triages them, and a critic pass lowers overconfident scores. A router sends the pick to Muse Code (Muse Spark 1.3), which writes a failing test first and then the smallest fix, under a per-issue USD cap. Plain Python, not a model, verifies Fail-to-Pass and runs the project’s full suite for regressions, with third-party code isolated from our keys. Every result is a proof bundle plus a hash-chained trace anyone can replay without API keys. The human only confirms paid runs and any upstream contribution.

Results: four real fixes VERIFIED with 0 regressions. aerkalov/ebooklib#359 came from a fully autonomous run started with one click: Spark Open chose the searches, picked the issue out of 25, and Muse Code hardened its path-traversal guard ($0.76). QWED-AI/qwed-verification#535: a taint-analysis bypass, full suite of 5,466 tests green ($1.25). ansible-runner#1548 ($1.23) and keynote-parser#91 ($0.013). All four were rebuilt from their committed bundles alone on GitHub Actions Linux, with no Muse and no keys, and came back VERIFIED. Failures stay on record: VulnClaw#310 was fixed correctly but shifted lines that its security allowlist tracks by line number, so ISSA AS reported REGRESSION instead of editing the allowlist. Total Meta spend for the build and all runs: about $8 of a $150 budget.

## More from Meta Global AI Developer Hackathon

- [Vary_artifacts](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/994)
- [ax0080](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/995)
- [MarketMate](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/996)
- [GaugeLens AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/998)
- [Pahami](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/999)
- [Maxim Zaks](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1000)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/997. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt