Linux Landlock - Nono Docs

nono uses Landlock LSM (Linux Security Module) on Linux to enforce capability restrictions at the kernel level.

What is Landlock?

Landlock is a Linux security module that allows unprivileged processes to sandbox themselves. Unlike traditional LSMs (SELinux, AppArmor) that require root configuration, Landlock can be used by any process to restrict its own capabilities. Landlock was merged into Linux kernel 5.13 (2021) and has been enhanced in subsequent releases.

How nono Uses Landlock

nono uses the landlock Rust crate to:

Detect the available Landlock ABI version
Create a ruleset with the allowed operations
Add path rules for each granted capability
Apply the ruleset to the current process

// Simplified: what nono does internally
let ruleset = Ruleset::new()
    .handle_access(AccessFs::from_all(abi))?
    .create()?;

ruleset.add_rule(PathBeneath::new(PathFd::new(path)?, access))?;
// ... add more rules ...

ruleset.restrict_self()?;
// After this call, restrictions are permanent

ABI Versions

Landlock capabilities have evolved across kernel versions:

Kernel	ABI	New Capabilities
5.13+	v1	Basic filesystem access control
5.19+	v2	`REFER` - rename/link across directories
6.2+	v3	`TRUNCATE` - file truncation
6.7+	v4	TCP `bind` and `connect` filtering
6.10+	v5	`IOCTL_DEV` - device ioctl filtering
6.12+	v6	Signal and abstract UNIX socket scoping

nono automatically detects the highest available ABI and uses it. On older kernels, some features are unavailable but core filesystem sandboxing still works.

Access Rights Mapping

nono maps its capability flags to Landlock access rights:

Read Access (`--read`)

AccessFs::ReadFile
AccessFs::ReadDir
AccessFs::Execute

Write Access (`--write`)

AccessFs::WriteFile
AccessFs::MakeChar
AccessFs::MakeDir
AccessFs::MakeReg
AccessFs::MakeSock
AccessFs::MakeFifo
AccessFs::MakeBlock
AccessFs::MakeSym
AccessFs::RemoveFile  // File deletion (needed for atomic writes)
AccessFs::RemoveDir   // Directory deletion (needed for rename() on directories)
AccessFs::Refer       // Rename/hard link across directories (ABI v2+)
AccessFs::Truncate    // ABI v3+

Note: RemoveFile and RemoveDir are both included because rename() requires removal rights on the source. This enables atomic write patterns (write to .tmp then rename to target) for both files and directories.

Full Access (`--allow`)

Both read and write access rights combined.

Network Filtering

Landlock ABI v4 (kernel 6.7+) added TCP network filtering:

AccessNet::BindTcp   // Control which ports can be bound
AccessNet::ConnectTcp // Control outbound connections

Network filtering requires kernel 6.7+. On older kernels, nono cannot enforce network restrictions via Landlock and will warn you.

Signal and IPC Scoping

Landlock ABI v6 added scope restrictions for process signals and abstract UNIX sockets:

Scope::Signal
Scope::AbstractUnixSocket

nono derives these scopes from the effective capability set:

Profile / capability setting	Linux V6 behavior
`security.signal_mode: "allow_same_sandbox"`	Requests signal scoping and fails closed if the kernel cannot enforce it
`security.signal_mode: "isolated"`	Requests signal scoping when available
`security.signal_mode: "allow_all"`	Does not request signal scoping
`security.ipc_mode: "shared_memory_only"`	Requests abstract UNIX socket scoping when available
`security.ipc_mode: "full"`	Does not request abstract UNIX socket scoping

shared_memory_only is the default IPC mode. On V6 kernels it prevents connecting to abstract UNIX sockets created outside the current Landlock domain. Profiles that need broader runtime IPC compatibility can opt into ipc_mode: "full". Use verbose dry-run output to see what would be requested and enforced:

nono run --dry-run -v --profile claude-code -- claude

Use nono why for a focused scope query:

nono why --scope signal --profile claude-code
nono why --scope abstract-unix-socket --profile claude-code

Pathname Unix Socket Mediation

Landlock filesystem rules and pathname Unix socket IPC are separate security surfaces. On Linux, a process that can reach a filesystem-backed Unix socket path may be able to communicate with a user-session service through that socket unless nono separately mediates AF_UNIX connect(2) and bind(2). For compatibility, this mediation is off by default on Landlock V4+ kernels. Profiles that need stricter IPC isolation can opt in:

{
  "linux": {
    "af_unix_mediation": "pathname"
  },
  "filesystem": {
    "unix_socket": ["$XDG_RUNTIME_DIR/bus"],
    "unix_socket_dir_bind": ["$TMPDIR/my-tool"]
  }
}

When enabled, pathname AF_UNIX sockets are default-deny and must match explicit filesystem.unix_socket* grants. Broad filesystem read/write grants no longer imply permission to talk to every socket under those paths. Allowed pathname sockets are resumed through Linux seccomp user notification. Because seccomp receives the syscall before the kernel copies sockaddr_un from userspace, a multi-threaded sandboxed process can race an allowed connect(2) or bind(2) by mutating that userspace address before the kernel continues the syscall. Denied sockets do not have this race because nono returns EACCES directly and the kernel never runs the original syscall. Treat this mode as a strong default-deny IPC hardening layer; avoid broad socket grants for untrusted multi-threaded workloads. High-risk socket categories include user service managers, session buses, keyring and credential agents, SSH/GPG agents, container runtimes, and package manager daemons. Hardened profiles should grant only the specific sockets they need.

If you run with linux.af_unix_mediation off in a public-facing, multi-tenant, privacy-sensitive, or otherwise high-risk deployment, run nono inside a stronger outer isolation boundary such as a MicroVM. Firecracker style MicroVMs are a good fit for service deployments where the VM boundary can isolate host user-session services and credentials from the sandboxed agent.

Enforcement Status

nono reports the enforcement status after applying the sandbox:

Status	Meaning
`FullyEnforced`	All requested restrictions are active
`PartiallyEnforced`	Some restrictions active, others unavailable (older kernel)
`NotEnforced`	Landlock not available on this system

Use -v to see the enforcement status:

nono run -v --allow-cwd -- command
# Output includes: Sandbox status: FullyEnforced

Checking Landlock Availability

# Check if Landlock is in the LSM list
cat /sys/kernel/security/lsm
# Should include "landlock"

# Check kernel version
uname -r
# 5.13+ required, 6.7+ for network filtering

Enabling Landlock

If Landlock is not listed in /sys/kernel/security/lsm, you may need to:

Check kernel config: Ensure CONFIG_SECURITY_LANDLOCK=y
Add to boot params: Add lsm=landlock,lockdown,yama,integrity,apparmor,bpf to kernel boot parameters
Reboot

Most modern distributions (Ubuntu 22.04+, Fedora 35+, Debian 12+) have Landlock enabled by default.

Irreversibility

Once restrict_self() is called:

The ruleset is permanently applied
No API exists to add more permissions
Child processes inherit the restrictions
The only escape is a kernel exploit

This is enforced by the kernel - nono cannot undo it even if it wanted to.

Debugging

If a command fails with permission errors:

Run with dry-run: See what capabilities would be granted
```
nono run --allow-cwd --dry-run -- command
```
Check verbose output:
```
nono run -vvv --allow-cwd -- command
```
Check dmesg for Landlock denials:
```
dmesg | grep -i landlock
```

Use strace: See which syscalls are being denied

strace -f nono run --allow-cwd -- command 2>&1 | grep EACCES

Limitations

Kernel Version Requirements

Basic sandboxing requires kernel 5.13+
Full filesystem control requires kernel 6.2+
Network filtering requires kernel 6.7+
Signal and abstract UNIX socket scoping requires Landlock ABI v6

No Network Filtering on Older Kernels

Without ABI v4 (kernel 6.7+), Landlock cannot filter TCP connections. nono will warn if you use --block-net on an older kernel. On kernels 6.7+, nono uses Landlock’s AccessNet::BindTcp and AccessNet::ConnectTcp to block TCP traffic. Note: DNS resolution (UDP) is not blocked by Landlock, only TCP connections.

Bind Mounts

Landlock follows bind mounts. If /home is bind-mounted to /mnt/home, access to one affects the other. This is usually not a concern but can be surprising in complex mount configurations.

Special Filesystems

Landlock may behave differently with special filesystems (procfs, sysfs, etc.). nono does not grant access to these by default.

Documentation Index

​What is Landlock?

​How nono Uses Landlock

​ABI Versions

​Access Rights Mapping

​Read Access (--read)

​Write Access (--write)

​Full Access (--allow)

​Network Filtering

​Signal and IPC Scoping

​Pathname Unix Socket Mediation

​Enforcement Status

​Checking Landlock Availability

​Enabling Landlock

​Irreversibility

​Debugging

​Limitations

​Kernel Version Requirements

​No Network Filtering on Older Kernels

​Bind Mounts

​Special Filesystems

​References

What is Landlock?

How nono Uses Landlock

ABI Versions

Access Rights Mapping

Read Access (`--read`)

Write Access (`--write`)

Full Access (`--allow`)

Network Filtering

Signal and IPC Scoping

Pathname Unix Socket Mediation

Enforcement Status

Checking Landlock Availability

Enabling Landlock

Irreversibility

Debugging

Limitations

Kernel Version Requirements

No Network Filtering on Older Kernels

Bind Mounts

Special Filesystems

References