Files
marqdo-cli/spec/security.md
T

303 lines
19 KiB
Markdown

# Marqdo Security Model (implemented status)
Status: **describes the M0.5 implementation, including honest gaps.** This is
not a security audit; it documents what the code actually enforces.
Normative implementation:
| Concern | Code |
|---|---|
| Sandbox modes, strict refusal | `crates/marqdo-sandbox/src/lib.rs` |
| Capability sets and checks | `crates/marqdo-host/src/capability.rs` |
| Host dispatch and per-op checks | `crates/marqdo-host/src/host.rs` |
| Project-rooted filesystem | `crates/marqdo-host/src/fs.rs` |
| Budgets | `crates/marqdo-host/src/budget.rs` |
| Effect tracing | `crates/marqdo-host/src/trace.rs` |
| Math/foreign adapters | `crates/marqdo-host/src/kernels.rs` |
| Foreign child processes | `crates/marqdo-foreign/src/lib.rs` |
| Plugin loading / in-process plugin calls | `crates/marqdo-host/src/host.rs` (`load_plugin`, `op_plugin_invoke`), `crates/marqdo-plugin-abi/src/loader.rs` |
| CLI flags and capability construction | `crates/cli/src/main.rs`, `crates/cli/src/commands.rs` |
## 1. Modes
`SandboxMode` has three values (`pure`, `project`, `strict`), parsed by
`SandboxMode::parse` and selectable on `run`/`call` with
`--mode <pure|project|strict>` (default: `project`).
| Mode | Intended guarantee | Actual status in this build |
|---|---|---|
| `pure` | No external I/O; memory computation and budgets only. | **Enforced.** `Capabilities::pure()` = `console`, `math` only; every FS/foreign/plugin grant flag makes the CLI exit 2 with "pure mode prohibits external I/O grants". |
| `project` | Operator-approved project resources under `CapabilityHost` + bounded resources. | **Enforced for the implemented operations.** Defaults equal `pure`; `--allow-read`, `--allow-write`, `--allow-foreign`, `--allow-plugin` add grants. `resource.open` supports stdin read, separate stdout/stderr write, granted project-file read/write (ADR-0025/0026/0040), granted in-process provider kinds (ADR-0033) and granted IP-literal TCP streams (ADR-0034); process/device remain unsupported. Sandboxing is cooperative (in-process checks), not OS isolation; in-process providers are trusted, not sandboxed. |
| `strict` | Untrusted documents under real OS isolation. | **Linux: enforced when the runtime probe passes (ADR-0021).** `resource.open` is **refused** in strict except for the existing `stdio/stdout` console channel (console grant); stdin/stderr, file/TCP/provider/process resources remain refused until a provider passes the isolation lane (CM-A11, ADR-0030/0034/0040). `PlatformCapabilities::detect()` runs a real Bubblewrap smoke test; success reports `Available { mechanism }`, failure reports `Unavailable { reason }`. Strict L1 supports the interpreter, console, native Math and the read-only bundled stdlib (published at `/run/marqdo-stdlib` for the Compat-C print chain); `--allow-read/--allow-write/--allow-foreign/--allow-plugin/--plugin/--trust-native-plugin/--kernel/--trace-file` are rejected with exit 2. macOS/Windows and Linux hosts without the mechanisms remain `Unavailable`; there is **no fallback** to an unrestricted process. |
`marqdo inspect <file> --platform` prints the detection report, for example:
```json
{
"os": "linux",
"arch": "x86_64",
"strict": {"state": "available",
"mechanism": "bubblewrap 0.11.1 (user/mount/PID/net/UTS/IPC namespaces + no-new-privs + rlimits)"},
"worker_processes": true
}
```
On a host that cannot create the namespaces the same block is
`{"state": "unavailable", "reason": "..."}` and strict refuses. Design rule
(blueprint §9): a cooperative in-process check is not a sandbox. See ADR-0021
for the Linux construction, L1 scope and non-goals; ADR-0019 still governs
every other platform.
## 2. Capability grants
`Capabilities` (`crates/marqdo-host/src/capability.rs`) is the effective grant
set for one run:
| Field | Default (`pure`/`project`) | Used by |
|---|---|---|
| `console: bool` | `true` | `resource.open("stdio","stdin",["read"])`, `resource.open("stdio","stdout"/"stderr",["write"])` and the stdlib console chain; a missing grant is `permission.denied` and stdout/stderr are also bounded by the output budget |
| `math: bool` | `true` | `math.*` host ops |
| `fs_read: Vec<String>` | empty | `fs.read` and project `resource.open("file", path, ["read"])` |
| `fs_write: Vec<String>` | empty | `fs.write` and project `resource.open("file", path, ["write"])` |
| `net_hosts: Vec<String>` | empty | exact IP-literal grants for `resource.open("tcp", "<ip>:<port>", ["read","write"])`; DNS is not resolved; `--allow-net <IP>` adds a grant (ADR-0034) |
| `process: bool` | `false` | required together with `foreign` for `foreign.run`/`foreign.call` |
| `foreign: bool` | `false` | foreign kernels |
| `plugins: Vec<String>` | empty | `plugin.invoke` and `StdHost::load_plugin` |
| `resource_providers: Vec<String>` | empty | registered in-process/worker/C-ABI `ResourceProvider` kinds; project mode only; helper processes and native libraries are trusted, not sandboxed (ADR-0033/0035/0036); local manifests are host-installed, unsigned (ADR-0037); the C worker helper is a trusted process boundary, not a sandbox (ADR-0038) |
Requests versus grants:
* `marqdo.yaml` `capabilities:` is a **request**. It is parsed and validated
(`crates/marqdo-package/src/lib.rs`) but is not consulted when building the
host capability set: only CLI flags/grant APIs produce grants. Declaring a
capability never grants it.
* `marqdo.yaml` `plugins:` likewise declares requirements only; it does not
install, load or authorize a plugin.
* CLI grant flags (`crates/cli/src/main.rs::SandboxArgs`):
* `--allow-read <GLOB>` and `--allow-write <GLOB>` (repeatable, project mode
only);
* `--allow-foreign` (project mode only; sets `foreign = true` and
`process = true`);
* `--allow-plugin <PACKAGE-NAME>` (repeatable, project mode only; matches the
manifest `plugin.name`, not the call namespace);
* `--plugin <PATH>` (repeatable, loads a plugin after the trust check);
* `--trust-native-plugin` (explicit in-process native trust decision).
* There is no network grant flag and no plugin-worker flag in this build.
### 2.1 Filesystem grants are project-relative globs
`fs.read`/`fs.write` paths go through `ProjectFs::resolve` **before** the
capability check:
* empty paths, absolute paths, `..` components and root/prefix components are
rejected with `permission.denied`;
* `.` components are dropped; the normalized relative string uses `/`;
* the capability patterns are matched with the `glob` crate against that
normalized relative path; a leading `./` is stripped on both sides, so
`--allow-read ./data/**` matches `data/a.csv` and does not match
`secrets/key`; `--allow-write output/**` does not grant reads;
* an invalid glob pattern is treated as non-matching (`Pattern::new(...).
unwrap_or(false)`), so it silently grants nothing (there is no diagnostic).
Additional FS limits (`crates/marqdo-host/src/fs.rs`):
* reads are capped at `MAX_READ_BYTES = 8 MiB`, must be regular files, must be
UTF-8, and the canonicalized path must stay inside the project root;
* writes create parent directories; before writing, the **closest existing
ancestor** is canonicalized and must stay inside the project root.
Known caveat: for writes, only the closest existing ancestor is checked and the
final write uses the non-canonical path. A symlink created between the check
and the write (TOCTOU) could escape the root. Reads canonicalize the final path
before reading. Full symlink/race hardening is an open P4 item.
### 2.2 Checks happen at execution points, not only at parse time
`Host::invoke` is the only path to effects. `StdHost::dispatch` routes every
`HostOp` and the checks run immediately before the effect:
* `fs.read` → `capabilities.check_read(path)` then `ProjectFs::read`;
* `fs.write` → `check_write` then `ProjectFs::write`;
* math ops → `capabilities.math` (in `crates/marqdo-host/src/kernels.rs`);
* foreign ops → `capabilities.check_foreign` (requires `foreign && process`);
* plugin invoke → the **complete export name** must be registered, then
`capabilities.check_plugin(package)` on **every call**; loading also checks
the same grant once;
* plugin host-capability requests → `CapabilityBroker` checks the package and
effective grants and **executes** granted `fs.read` through `ProjectFs`,
returning the real file bytes. Other core capabilities (`fs.write`,
`net.request`, `process`, console/math/log/...), and vendor capabilities
without a trusted executing handler, return `capability.unavailable`; they
never pretend an empty reply is success. Unknown/vendor names without a
trusted policy entry become `capability.unavailable`/`permission.denied`;
* print → output budget only.
Static analysis may warn, but it never replaces these runtime checks. There is
no host bypass for tests: tests use the same `StdHost`/`Capabilities` types.
## 3. Resource budgets
`Budget` (`crates/marqdo-host/src/budget.rs`) is per run:
| Limit | Default | Enforcement |
|---|---|---|
| evaluation fuel | `DEFAULT_FUEL = 5_000_000` | 1 unit per statement, 1 per expression node, 1 per branch condition, 1 per loop iteration; exhaustion → `resource.exhausted` |
| call stack depth | `DEFAULT_STACK_LIMIT = 256` | checked before every function call → `resource.exhausted` |
| program output bytes | `DEFAULT_OUTPUT_BYTES = 1 MiB` | `print`, `stdio/stdout` and `stdio/stderr` output plus `fs.read` contents are charged; overflow → `resource.exhausted` |
| wall clock | none by default | `--timeout <seconds>` sets a deadline; checked between statements/iterations and before loop conditions → `resource.exhausted`. It does **not** preempt a running host call. |
| addressable memory | 64 MiB committed total, 16 MiB per region, 64 KiB pages; `--max-memory` / `--max-region-memory` override | Runtime-owned `MemoryStore` (ADR-0022); create/grow checks both budgets before allocation and failed growth is atomic → `resource.exhausted`. Memory is pure computation and needs no capability grant in any mode. |
| resource streams | stdin reads over a shared run cursor; stdout/stderr writes charged to the existing output budget; file reads and writer buffers bounded by the 8 MiB per-file limit | Host-owned resource table (ADR-0025/0026/0040); output budget is checked before a stdout/stderr write is buffered; a failed check produces no bytes; writer buffers flush only on `close`. No separate I/O byte budget yet. |
CLI overrides (`run`/`call`): `--fuel`, `--stack`, `--max-output`,
`--timeout`. Additional fixed limits:
* foreign kernel: `StdHost.foreign_timeout = 10s` and
`foreign_output_limit = 1 MiB` (stdout+stderr combined) per call.
These are not configurable from the CLI in this build;
* plugin wire: request ≤ 16 MiB, result ≤ 64 MiB, error ≤ 1 MiB, metadata
≤ 64 KiB (see `include/marqdo_plugin_abi_v1.h`);
* math kernel: 1 MB source, 256 parse depth, 512 AST nodes, exponent ≤ 10,000,
exact-result ≤ 1,000,000 bits (`crates/marqdo-math`).
These are cooperative budgets in the interpreter and the supervisor. Only
foreign child processes are terminated by the host (kill on timeout/output
overflow). In-process math and native plugins are not OS-limited.
## 4. Effect tracing
`EffectTrace` records one entry per `Host::invoke` when enabled
(`--trace`, or implicitly `--json` for `run`/`call`). Entry fields:
```json
{"op":"print","module":"hello.mq.md","span":{"start":43,"end":57},
"duration_micros":3,"result":"ok"}
```
* `op` is the stable `HostOp::name()` (`print`, `math.eval`, `foreign.call`,
`fs.write`, `plugin.invoke`, ...);
* `module` is the initiating module id; `span` is the call site when known;
* `result` is `"ok"` or the failure code (for example
`permission.denied`);
* argument values and file contents are **not** recorded, so ordinary secrets
do not leak into traces by design. Paths are also not recorded today.
* In text mode trace entries are printed to stderr (`trace: ...`); in JSON
mode they are a `trace` array next to separate program `stdout`/`stderr`
strings. Program output is always separate from diagnostics/trace, and
`stdio/stderr` never shares the `stdout` buffer.
## 5. Foreign kernels: process isolation, not a sandbox
`crates/marqdo-foreign/src/lib.rs` runs Python/Node one-shot:
* the program is written to a private temp file (mode 0600 on Unix), the child
runs with a fresh private temp directory (mode 0700) as cwd, and both are
removed on drop;
* direct `argv` exec (no shell, no interpolation); stdin carries the versioned
request envelope; stdout/stderr are always piped and captured;
* the environment is cleared except `PATH`;
* on Unix the child is a process-group leader (`setpgid`) and the wall-clock
deadline/output overflow kills the whole group, including grandchildren; a
2-second reader grace prevents a surviving pipe holder from hanging the
host. Non-Unix targets only kill the direct child;
* a non-zero exit is data (`RunStatus::Failed`, stderr preserved); a timeout is
`TimedOut` for `run` and an error for the typed `call`.
What this is **not**: an OS sandbox. The child runs with the host user's
privileges, can open absolute paths, use the network and spawn further
processes. Clearing the environment and changing cwd are hygiene, not
confinement. Grandchildren are not killed when the direct child is killed.
Strict L1 rejects Foreign execution before any sandbox starts, so foreign
execution is only reachable in `project` mode with `--allow-foreign`. L2
inheritance is future work (ADR-0021).
## 6. Plugins: in-process native trust boundary
* The only plugin transport implemented in this build is **in-process native
loading** (`libloading`) after `manifest.validate()`, exact target-triple
artifact resolution, sha256 verification, ABI negotiation and export
metadata cross-checking (`crates/marqdo-plugin-abi/src/loader.rs`).
* Loading requires **three** independent things:
1. `--plugin <manifest-or-dir>` names the plugin;
2. `--trust-native-plugin` is an explicit unsafe trust decision;
3. the manifest package name is granted with `--allow-plugin <name>` (checked
again on every invocation).
Without trust, `StdHost::load_plugin` returns `permission.denied` with the
hint that no isolated worker transport exists. There is no isolated worker,
Wasm component or out-of-process plugin path in this build.
* A native plugin runs in the host address space. It can read host memory,
call arbitrary syscalls, bypass the capability model entirely and crash or
corrupt the host process. **A C ABI is not a sandbox.** `--trust-native-plugin`
must be treated as equivalent to running the plugin's code as the host user.
* Every export is registered by its complete dotted name; a plugin may export
several prefixes, and a conflicting name rejects the whole load atomically.
`check --plugin <manifest>` (or `run/call --plugin`) validates export
signatures statically without loading a shared library.
* Plugin call sites are written as `<export>(...)` in a function
body (for example `solver.add(a=20, b=22)`). The namespace is derived from
the **first** manifest export prefix. Calls are made through
`HostOp::PluginInvoke` with a typed `wire_version` 1 request; the plugin's
typed response becomes a Marqdo value (`Unit` when absent). Named arguments
are required in this build; positional plugin arguments are `runtime.type`.
* Plugin errors map to structured runtime errors: ABI → `plugin.abi_mismatch`,
capability → `permission.denied`, wire/value → `runtime.type`, unsupported →
`kernel.unavailable`, everything else → `plugin.failure`. An unloaded
namespace is `symbol.unknown`.
* The plugin's host-capability requests are denied by default: the CLI loads
plugins with `LoadOptions::default()`, i.e. **no capability handler**, so
every `request_capability` call fails. The `requires:` list in `plugin.yaml`
is only a declaration.
* `marqdo plugin inspect` never loads code: it validates the manifest and
verifies the artifact hash only.
* Installation, inspection, loading and authorization are four separate
decisions. Copying a plugin into the project or declaring it in
`marqdo.yaml` grants nothing.
Known limitations: single namespace per plugin (derived from the first export),
no hot unload/reload, no plugin capability bridge from this CLI, native crash
is a host crash, `plugin.invoke` cannot be used in `pure`/`strict` mode.
## 7. `check` and `inspect` are side-effect free
Neither command executes project code. Specifically:
* `check` parses, indexes, maps to IR, runs static call validation and reports
diagnostics. It never calls a function, never spawns a foreign kernel, never
loads a native plugin and never writes files.
* `inspect` with `--ast`, `--units`, `--signature`, `--ir`, `--refs` or
`--platform` does the same static work and prints metadata. `--refs` may
**read** other project source files on demand to resolve link targets, but
only parses/analyzes them; nothing is executed or written.
* `plugin inspect` reads `plugin.yaml`, validates it, resolves the artifact for
the host target and computes/compares its sha256. It does not `dlopen`,
instantiate or call the plugin (`"loaded": false` in JSON output).
* `run`/`call` first perform the same static load; if any diagnostic has error
severity, execution is refused before the host or any kernel is touched.
* `fmt` only normalizes line endings/trailing newlines and writes the document
when not in `--check` mode; it does not execute anything.
## 8. Open security items
1. Strict L1 is implemented and VM-verified on Linux (ADR-0021); strict L2
(Worker/Foreign inheritance) is not implemented, so those surfaces are
refused in strict.
2. `console` capability is not enforced; print is always allowed.
3. Network capability exists in the model but no host operation uses it; there
is no `--allow-net` flag.
4. Write-path symlink TOCTOU (see §2.1).
5. Plugin **worker crash isolation** and the worker capability bridge are
implemented (not usable in strict L1): both native and worker plugins can
reach the broker's
`fs.read` execution when granted. Other core capabilities and vendor
handlers remain unavailable. Broker execution is a cooperative API
contract, not an OS sandbox; a worker and a native plugin can still issue
direct syscalls.
6. `plugin_handle` wire values are structurally reconstructable, but host
authority comes only from the `PluginHandleTable` owner/run/generation
check; untrusted plugin/foreign output containing a handle is rejected
(ADR-0014, `value-wire.md` §4.1).
7. Budgets do not cover all allocation paths (for example list/table growth
inside the interpreter is bounded only by fuel and memory).
8. `--json` run output includes the full program stdout string; that is the
program's own output, not a host secret, but it is embedded in the JSON
document consumers must parse.