What happened
CVE-2026-95183 (unrated, published 2026-10-16 per NVD REST) — an issue in the context-compaction feature of DeepSeek Harness (dsh) commit c291e796 allows attackers to execute a prompt injection attack via crafted external content.
Why it matters
Evaluation harnesses are supposed to glove the model; a prompt-injection reachable through the harness's own context-compaction means an attacker who controls external content entering an eval can steer the model under test — undermining the integrity of benchmark/capability results in AI-red-team and eval infrastructure.
Attack vector
Crafted external content injected through the context-compaction feature of the harness executes a prompt injection attack against the evaluated model/session.
Affected systems
DeepSeek Harness (dsh) at commit c291e796
Mitigation
No public advisory or patch reference yet (CVE published 2026-10-16, unrated); pin/update dsh and treat untrusted external content that flows into context compaction as tainted.