The vulnerability resides in the run_code function of the Code Mode Sandbox component of deepseek-harness. An attacker can trigger a sandbox escape that undermines the containment guarantees provided by the worker. While the vendor has stated that the worker is meant for containment rather than enforcing a security boundary, the manipulation defeats this containment, potentially allowing the attacker to gain elevated privileges or execute code outside the intended sandbox. The impact is therefore a loss of isolation rather than a full remote code execution, but it still permits privilege escalation or access to sensitive resources within the host environment.

deepseek-ai deepseek-harness, version 0.1.0-rc.7 and earlier are affected.

The CVSS score of 5.3 categorizes this vulnerability as moderate severity. No EPSS score is available, and the issue is not listed in CISA KEV, indicating that there is no publicly documented exploitation activity yet. The attacker can remotely interact with the run_code endpoint to trigger the sandbox issue, as the vulnerability is accessible over the network. The lack of an official patch or workaround means that affected deployments remain susceptible until the vendor addresses the containment weakness.