Skip to content

Security: andreibesleaga/agentic-system-core

SECURITY.md

Security policy

Reporting a vulnerability

Report it privately, through this repository's private vulnerability reporting: open the repository's Security tab and choose Report a vulnerability. The report goes to the maintainer and is not public while it is being handled.

Please do not open a public issue for a security problem, and please do not post the details anywhere public until a fix is released.

If private reporting is unavailable to you, use the contact page at https://andreibesleaga.com/contact/ and ask for a private channel; do not put the details in the first message. Both routes, in this order, are the Contact: lines of this project's published security.txt, at https://agenticsystemcore.com/.well-known/security.txt (RFC 9116).

Useful in a report: what you did, what happened, what you expected, the version or commit, and — if the finding is about a published site — the exact address.

What to expect

This project is maintained by one person, in their own time. There is no service level, no paid support and no bug bounty: no payment is offered or made for a report. Reports are read and answered as soon as the maintainer can, and a reporter who asks to be credited is credited in the release notes.

Supported versions

Version Supported
1.1.x yes (the latest patch)
1.0.x no — upgrade to the latest 1.1.x (the change log says how)
any 1.0.0-rc.* no — upgrade to the latest 1.1.x
anything older no

Fixes are made on the current line only, in a new release; there is no long-term-support branch, and no fix is made to a release candidate. Every released version stays readable at its tag in this repository and its packages stay installable, but none of them is supported once a later version is out.

Coordinated disclosure

A fix is released first, and the advisory follows it. While a report is being handled, its details stay between the reporter and the maintainer. Once the fixed release is out, the maintainer publishes a GitHub security advisory for this repository that names the affected and the fixed versions and credits the reporter who asked to be credited; a GitHub advisory can carry a CVE identifier. Please keep the details private until that advisory is published.

What this project is, and what that means for a report

The engine is a command-line program that reads files and writes files. It has no database, its verbs open no network listener, and the published sites are static files. Its network port refuses every call at 1.0 and the release runs offline, so the main attack surface is the files a Bundle is built from. Two programs in the package touch a network, each only when a person runs it (and, from 1.1, one form of a verb listens on the machine's own loopback address, below):

  • agsc-host serve serves a build directory read-only over plain HTTP. It binds localhost:8080 unless told otherwise (--bind, --port), answers GET and HEAD only, and never follows a link out of the directory it serves; any origin other than loopback needs a TLS terminator in front. It answers only a request whose Host header names localhost, a loopback address or the address it is bound to, and refuses any other with 421, so a web page that points a name of its own at the machine cannot read what it serves; it never serves a file or folder whose name begins with a dot, except .well-known.
  • tools/validate-wellknown <URL> fetches the discovery document it is given and its targets: https only, a closed list of refused addresses checked again on the open socket, at most three redirects, a deadline and a byte cap per response (AGSC-11-07 to AGSC-11-10).
  • From 1.1, agsc mcp --http <port> serves the ten tools over Streamable HTTP (AGSC-09-18). It binds a loopback address only (127.0.0.1, or with --bind another address in 127.0.0.0/8 or ::1; any other is refused before anything starts), writes no file and sends nothing. It refuses with 403 a request whose Origin or Host names any host but 127.0.0.1, localhost, [::1] and the bound address, so a web page of another site, or a name of its own pointed at the machine, cannot call a tool; it serves /mcp alone and POST alone. The checks and the protocol are those of the official MCP SDK (@modelcontextprotocol/server, @modelcontextprotocol/node).

The interesting reports are therefore usually about input handling: a crafted item, attachment, diagram source, imported corpus or peer document that makes the engine write something it should not, escape the Bundle root, or produce a surface that misleads a reader or an agent.

One input is trusted rather than checked: the Bundle's own .git directory. The engine runs git in the Bundle to read its history, and git reads that directory's configuration. The engine switches off the two settings known to start a program during those reads (core.fsmonitor and log.showSignature), but a repository that came from someone else, for example unpacked from an archive, can carry other configuration. Clone a Bundle you did not make, or inspect its .git/config, before you run the engine in it.

The threat model, and the honest limits of what the checks prove, are written down in docs/SECURITY-CONSIDERATIONS.md. The specification's own rule on the subject (spec/08-governance.md, AGSC-08-19) says of the four agent-safety lints:

These lints prove neither safety nor the absence of novel injection; hashes and attestations prove only that an artefact is what was published. An implementation MUST NOT claim more.

That sentence is the standard this project holds itself to. A report that shows the limit is narrower than claimed is as welcome as a report of a defect.

There aren't any published security advisories