πŸ‡«πŸ‡· Version franΓ§aise

Security Policy

Supported Versions

Version Supported
Latest stable release βœ…
Older releases ❌

Only the latest stable release receives security fixes.

Reporting a Vulnerability

Please do not open a public issue for security vulnerabilities.

Report vulnerabilities through GitHub Private Vulnerability Reporting on this repository ("Security" tab β†’ "Report a vulnerability").

What to include:

  • A description of the vulnerability and its impact.
  • Steps to reproduce (proof-of-concept code or configuration if possible).
  • The affected component (tool, LLM provider, memory provider, orchestrator…).

We aim to acknowledge reports within 72 hours and to provide a remediation plan or fix within 30 days for confirmed issues.

Threat Model β€” LLM-Driven Tool Execution

Orkeon is an AI-agent orchestration framework: LLM output can trigger tool execution. This makes certain components security-sensitive by design, and you should treat any crew configuration as part of your attack surface:

  • Code/shell execution tools (ShellCommandTool, SecureCodeInterpreterTool): shell commands are allowlist-restricted by default (read-only set) and interpreters (node, dotnet, npm, find) plus mutating git subcommands require an explicit opt-in. shell_command is confined by the operating system by default β€” bubblewrap on Linux, Seatbelt on macOS, an AppContainer inside a Job Object on Windows: the agent's mounts, the system directories read-only, no network β€” and fails closed when the backend cannot run; it is unconfined only when the host runs in a container (the container is the boundary) or with Orkeon:Tools:Shell:Confinement:Mode = None (ADR-014). code_interpreter runs in the Docker sandbox (DockerSandbox) and fails closed: when Docker is unavailable it refuses to fall back to plain host execution unless Orkeon:CodeSandbox:AllowHostExecution is set.
  • Network tools (WebScrapeTool, HttpApiTool, web search, rag_ingest and the opt-in RAG web fallback): URL validation is fail-closed β€” private, loopback, link-local, IPv6-unspecified, multicast, NAT64 and cloud-metadata addresses are denied by default (SSRF protection), and URL ingestion refuses to fetch at all when no IUrlValidator is registered (AddOrkeonInfrastructure registers one). Redirects are not followed: AddOrkeonWebTools gives the web tools their own named HttpClient and the RAG page loader its own typed client, both with AllowAutoRedirect = false, so a redirect cannot carry a request past the validator that cleared the first hop.
  • File system tools: all file access goes through the Virtual File System (IFileSystemService), validated against configured mounts and access rights.
  • Database tools: queries pass through IDatabaseSecurityPolicy (statement-type allowlist).
  • E-mail tools (email_*): they act on a real mailbox and read mail written by strangers. An agent only names an account; servers, credentials and each account's mandatory Rights are operator configuration, and a secret is only ever the name of an environment variable. Sending fails closed β€” Send:AllowedRecipients empty means nobody, and the SMTP envelope is the checked recipient list, never read from headers. Every message read is marked untrusted and screened for prompt injection, which flags but can be evaded. Every account is visible to every crew and .ork.ts script that resolves the same settings file (crews hosted by orkeon-host share the host's): declare accounts in the crew's own settings file β€” not in environment variables or an appsettings.json of the directory runs start from, which the .NET host reads for every run β€” grant the fewest rights, prefer email_draft to email_send, and never give an untrusted crew both e-mail Read and an outbound channel (http_api, the web tools). OAuth tokens sit in a file under the per-user settings directory, out of reach of the VFS tools but not of a shell or code tool running as the same OS user. Details: E-mail tools and Security.
  • Drive tools (drive_*): they act on a real Google Drive or OneDrive and read files anyone may have shared. An agent only names an account; the OAuth client, each account's mandatory Rights and its Root folder are operator configuration, and an item outside the root is reported as not found. Sharing fails closed β€” an empty Share:AllowedRecipients allows nobody, and public links stay off until Share:AllowPublicLinks is set. drive_read marks what it reads untrusted and screens it for prompt injection. The bearer token is sent to the provider's API host only β€” never to a download link, an upload session or a paging link elsewhere. As for e-mail, declare accounts in the crew's own settings file, grant the fewest rights, and never give an untrusted crew both drive Read and an outbound channel. Details: Drive tools and Security.
  • Prompt injection: any content fetched by tools (web pages, files, database rows) may contain adversarial instructions. Run agents with the least-privileged toolset your use case allows.

Vulnerabilities in these layers (allowlist bypass, SSRF filter bypass, VFS escape, sandbox escape, secret leakage in logs, a send outside an account's recipient allow-list) are considered high severity β€” please report them privately.

Hardening Recommendations

  • code_interpreter is never exposed as an IBaseTool: SecureCodeInterpreterTool is registered as a concrete type only, so no agent can call it unless you expose it yourself.
  • shell_command is different β€” it ships registered. AddOrkeonCodeTools() registers it as an IBaseTool, and the shipped runners (orkeon run, orkeon-repl, orkeon-host) call that method unconditionally, so the tool is in the catalogue out of the box. Its default allowlist is read-only (ls, cat, pwd, grep… plus read-only git subcommands); interpreters and mutating git stay off unless you set Orkeon:Tools:Shell:AllowInterpreters, which is RCE-equivalent and makes the tool emit a security warning. Prefer adding the one command you need to Orkeon:Tools:Shell:ExtraAllowedCommands; Orkeon:Tools:Shell:AllowedCommands replaces the allowlist outright. To keep shell_command out of an agent's reach entirely, compose your host without AddOrkeonCodeTools().
  • Use DockerSandbox for any code-execution scenario with untrusted input, and leave Orkeon:CodeSandbox:AllowHostExecution off.
  • Configure API keys via environment variables or a secret manager β€” never in YAML crew definitions.
  • Enable memory encryption at rest (EncryptedMemoryProviderDecorator) for sensitive workloads.

Verifying What You Install

Every artefact β€” NuGet packages, installers, CLI archives, .deb, MSIs β€” is built by a public GitHub Actions workflow and carries a GitHub-signed build provenance attestation (SLSA v1) naming the workflow, the tag and the commit that produced those exact bytes. NuGet.org publishing goes through Trusted Publishing (OIDC): no long-lived API key exists. Every release also ships SHA256SUMS (and SHA256SUMS.msi for the MSIs) and β€” for releases cut since the SBOM step landed in September 2026 β€” a CycloneDX SBOM covered by the same attestation. Two things carry no attestation: the dev builds of main (<version>.dev.<run number>, GitHub Packages only) and the ghcr.io/orkeon/orkeon container image on GHCR.

gh attestation verify orkeon-cli-<version>-osx-arm64.tar.gz --repo Orkeon/orkeon

A package downloaded from nuget.org is repository-signed by nuget.org, which changes its bytes: strip that signature first (scripts/nupkg-unsign.py), then verify. The exact commands, what each mechanism proves and does not, and what to do when a check fails are in Verify what you install. An artefact that fails verification is a security report, not a support question.

apt archive signing key

The Debian / Ubuntu repository (Install with apt) is signed with a dedicated OpenPGP key, Orkeon Archive Signing Key <arion@orkeon.org>. The first download of that key is trust on first use, so its fingerprint and the SHA-256 of the binary keyring file the apt branch serves (orkeon-archive-keyring.gpg) are published here, on main:

orkeon-archive-keyring fingerprint: 4765 9C57 4882 5078 5C78  2C37 1CF8 CD8A 4C20 1A2E
orkeon-archive-keyring.gpg sha256: 0cc5e804e1ee49c50ec5b23145ee7ee65d4834a08b8dc38914674249dfcc6a21

The CI guard checks these two lines against the certificate committed in installers/apt/.

The public certificate is versioned, ASCII-armoured, as installers/apt/orkeon-archive-keyring.asc; the binary keyring is that file passed through gpg --dearmor. A CI guard fails when the values above disagree with it, or with the installation page.

What apt checks by itself, once the source is declared with Signed-By:: the signature of each channel's index (InRelease) against this key only, then the SHA-256 of every package against that index. You check the keyring's SHA-256 once, when you add the source β€” the installation block does it.

Key policy.

  • An OpenPGP v4 key: an ed25519 primary key used for certification only and kept offline, and an ed25519 signing subkey valid two years β€” the only secret the publishing workflow holds. Index signatures use SHA-256.
  • The signing subkey is extended at least six months before it expires; CI fails when its expiry is less than 180 days away. An expired key makes apt update fail everywhere (EXPKEYSIG).
  • A new subkey reaches users through the orkeon-archive-keyring package, by apt upgrade, one release before the repository is signed with it: nobody has to act.
  • A revocation β€” a compromised key β€” is announced here and on the installation page, with the new fingerprint and keyring SHA-256. It is the only case where users must download a new key themselves. Publication of the repository stops until the new key is in place.