Skip to content

Security: buttercode101/agentscope

Security

SECURITY.md

Security Policy

Reporting a vulnerability

If you discover a security issue in agentscope, please report it privately rather than opening a public issue:

Include steps to reproduce, the impact you believe it has, and any suggested fix. We acknowledge reports within 48 hours and aim to resolve confirmed issues promptly. Please do not disclose the issue publicly until a fix is available.

Data handling

generate.py reads local agent state (cron job definitions, execution logs, session records — by default under ~/.hermes/, or paths you configure) and renders a self-contained static HTML dashboard.

  • Data never leaves the machine. The generator makes no network calls, runs no server, and uploads nothing. The public website is a static product page; your generated dashboards are local files.
  • Read-only supervision. agentscope inspects state; it does not control agents, and the dashboard contains no operation buttons.
  • Sensitive local data (e.g. prompts, credentials in logs) can end up in the generated HTML. Treat the output file like the source data: do not publish it to a public host without review.
  • Local helper files. Sample data under data/ is illustrative only; do not put real credentials into scanned paths.

Known limitations (not vulnerabilities)

  • Generated HTML is not sanitised for every possible log format; source data containing script-like strings is rendered as evidence, not executed by this tool — but a user who opens a maliciously crafted source file could be exposed. Do not scan untrusted inputs in a browser you would not trust with the source itself.
  • agentscope reports system-recorded state; it cannot verify things that never happened. "Running" means observed evidence, not proof of external truth.

Scope

In scope: generate.py and the published demo page.

Out of scope: hosted AI-agent platforms (Hermes Agent, LangSmith, and similar) except as data sources, and third-party hosts (Vercel, GitHub) except as they deliver this repository's artifacts.

P0 note — commission evidence removed from public repo

Commission evidence (the Madlanga commission evidence API previously mixed into this dashboard repo: api/, data/commission.json, evidence.db, data/snapshots/*.bin, data/*.bin) has been moved out of this public repository and removed from the current HEAD.

Residual risk: earlier commits still contain this evidence. Removing files from HEAD does not purge git history. A full history purge would require a force-push, which was not performed as part of this change. Anyone with access to an old commit, mirror, fork, or GitHub cache can still retrieve the previously exposed files. Treat previously public evidence as exposed, rotate/revoke anything that was secret in it, and consider a history purge (e.g. git filter-repo

  • force-push) as a separate, explicitly approved decision.

There aren't any published security advisories