Unattended AI Agent Framework Patches Flaws Across 2.1M-Star Repos

Aeon is an MIT-licensed agent framework whose entire runtime is a GitHub repo plus Actions cron jobs — no server, no daemon. Skills live as single Markdown files and state is committed directly to git, letting the agent operate on a schedule with no human in the loop.

Its flagship production use case is an end-to-end vulnerability disclosure pipeline: a scanner skill audits open-source repos, writes patches, opens PRs or private security advisories, and tracks fixes through to merge. The public disclosure board currently lists 70 hardened repositories with 2.1 million combined GitHub stars, including High- and Critical-severity fixes in projects from Alibaba, xai-org, Tencent, and Vercel.

Other live use cases are similarly ambitious: skills rewrite their own code by generating and scoring four competing versions before shipping the winner as a PR; self-replicating agent fleets run with built-in cost accounting and anomaly alerts; and on-chain USDC payroll for contributors is secured with idempotency keys and a two-phase execute flow to prevent double-payment.

The real lesson for engineers isn't about model capability — it's the guardrails: read-only skills, daily caps on repair PRs, a ban on direct commits to main, and backpressure rules that stop the agent from overwhelming its own maintainers. Unattended autonomy here was earned by engineering the brakes first.