The defect lies in the context‑compaction routine of DeepSeek Harness, engineered in commit c291e796. Its purpose of condensing user prompts inadvertently permits arbitrary code to be embedded within the generated prompt. The vulnerability results in a classic prompt injection flaw, enabling an attacker to cause the harness to execute malicious code under its own process. This weakness is reflected in CWE‑94 (Improper Control of Generation of Code).

Any deployment of DeepSeek Harness that incorporates the context‑compaction logic from the specified commit is vulnerable. The vendor has not catalogued a specific release number for the affected build, thus installations that predate the patch or have not yet updated past this commit are at risk.

The exploit is remotely achievable through externally supplied content, with no authentication required. The CVSS score of 8.8 indicates high severity. The EPSS score lies below 1 %, suggesting a low current probability of real‑world exploitation, and the issue is not listed in the CISA KEV catalog. Nonetheless, because it permits arbitrary code execution by manipulating prompt content, the flaw remains a serious risk, especially if the context‑compaction feature is exposed to untrusted inputs. Based on the description, it is inferred that the attack vector involves an attacker providing crafted content that is processed by the compaction component before being used to generate a prompt.