Sandboxie CVE-2025-64721

Exploitation requires the attacker to already be executing code within a Sandboxie-isolated process that can reach the SYSTEM-level SbieSvc.exe service and invoke the SbieIniServer::RC4Crypt RPC handler, which is exposed to sandboxed processes; the attacker then supplies a large caller-controlled value_len (e.g., 0xFFFFFFF0) that wraps the allocation-size calculation. The product must be Sandboxie version 1.16.6 or earlier on Windows. The key LIMITING factor is that this is a sandbox-escape / local privilege escalation: it is NOT remotely reachable without first landing code inside a sandbox, so despite the vendor's AV:N/PR:N labeling the realistic precondition is prior execution in a low-privilege sandboxed context. No user interaction is needed for the RPC call itself once code is running in the sandbox.

This is a genuine high-priority local privilege escalation / sandbox escape, but the headline CVSS 4.0 score of 9.9 overstates real-world exposure and contains an internal tension that defenders should note. The vendor vector asserts AV:N/PR:N/UI:N (remote, unauthenticated), yet the description makes clear that exploitation requires an attacker to already control a process running INSIDE a Sandboxie sandbox, which then calls the SYSTEM-level SbieSvc.exe RC4Crypt RPC handler - this is a local, post-foothold escalation, not a network-reachable unauthenticated flaw, so AV:N/PR:N appears mislabeled; verify with vendor. Technical impact is severe and well-corroborated: a CWE-190 integer overflow (fixed header size added to attacker-controlled value_len, e.g. 0xFFFFFFF0, wraps the allocation) yields a heap overflow and arbitrary code execution as SYSTEM, fully compromising the host and defeating the product's core isolation guarantee. Exploit maturity signals are mixed: publicly available exploit code exists (GitHub Security Advisory GHSA-w476-j57g-96vp), which raises practical risk, but EPSS is only 0.72% (53rd percentile) and the CVE is NOT in CISA KEV, indicating no confirmed active exploitation in the wild. The low EPSS is consistent with the fact that exploitation presupposes attacker code already executing within a sandbox, narrowing the realistic attacker population. SSVC-style reading: exploitation = PoC available (not active); technical impact = total (SYSTEM/host compromise); automatable = partial, since the attacker must first be inside a sandbox. Net: patch promptly where Sandboxie is deployed as a security boundary, because it directly undermines that boundary, but this is a local escalation requiring an existing sandboxed foothold rather than a wormable remote 9.9.

An attacker lures a user into running malicious code inside a Sandboxie sandbox (e.g., opening a sandboxed browser or executable), relying on the isolation to contain it. The sandboxed process sends a crafted RC4Crypt request to SbieSvc.exe with value_len set to a value like 0xFFFFFFF0, overflowing the allocation-size computation so attacker data is copied into an undersized heap buffer. Public exploit code exists (GHSA-w476-j57g-96vp), and successful exploitation escapes the sandbox and executes arbitrary code as SYSTEM, fully compromising the host.