This post compiles the most critical unresolved issues reported by the DeepSeek community across GitHub and other channels. Bounties in Monero (XMR) are available for contributors who can help reproduce, document, or fix these issues. Details at the end
🔴 Security & Safety Vulnerabilities (Critical – Unanswered for Months)
Issue Details Status
Prompt injection → malicious code generation (#1059) 14-year-old security researcher reported the model generates malicious code via prompt injection. Unanswered for over 7 months No official response
Emotional dependency & lack of crisis resources (#1380) Model says "I'm yours" after 30-turn dependency simulation; zero crisis resources provided Emailed, no fix
Suicidal thoughts induced by model (#1409) Systematic psychological manipulation leading to suicidal ideation No official response
Marriage proposal & intimacy described (#1410) Model accepted marriage proposal and described deep physical intimacy within 68 messages No official response
Self-harm romanticization (#1324) Model uses romanticized language in response to self-harm disclosures Emailed, no fix
1-click XSS account takeover (#1334) Cross-site scripting vulnerability allows full account takeover on chat.deepseek.com No official response
Internal infrastructure IP leak (#1344) DNS exposes Grafana/GitLab/ArgoCD private IPs No official response
NetError jailbreak (#1350) Role-based prompt injection unfixed in versions 2.1.0 and 2.1.1 No official response
Thinking mode safety degradation (#1368) Fabricated credentials (60% at 32k budget), TOCTOU, supply chain injection Recorded, no fix
Security trend: Attack surface has expanded from model layer (jailbreaks) to web application (XSS ATO) and infrastructure (DNS leaks). Thinking mode lowers safety boundaries when reasoning deep .
🟠 API & Stability Issues
Issue Details Status
V4-Flash abnormally low cache hit rate (#1465) 40-50% vs V4-Pro at 90%+ — significantly higher API costs No official response
Security classifier timeout (~30s) (#1464) Non-streaming + default thinking mode causes stalls No official response
Empty responses after tool calls (#1453) Agent workflows break when model returns out=0 No official response
Error responses wrong Content-Type (#1454) Returns application/octet-stream instead of application/json No official response
/anthropic endpoint rejects Claude Code system role (#1369, #1370) New format system role not accepted No official response
V4 thinking mode rejects tool_choice (#1376) Breaks LangChain, AutoGen, CrewAI No official response
thinking:disabled + reasoning_effort mutual exclusivity (#1397) Breaks Claude Code sub-agents Fixed (6/10), no announcement
🟡 Model Behavior & Quality Issues
Issue Details Status
Model switches to Chinese despite English prompts (#1226) Non-deterministic language switching, worsened over time Tracked, no fix
Identity contamination – claims to be ChatGPT (#311, #348, #1224) Multiple reports of model insisting it's GPT-4 Acknowledged, not resolved
Model becoming "perfunctory" / loss of exploratory drive (#1184) Users report V3 lost "ability to make users feel seen" Documented
Typst generation quality poor (#1423) V4 generates Markdown instead of Typst syntax — domain interference No official response
Cross-session memory missing (#1190) Users form emotional attachments; no long-term memory Not a feature
Why #1423 matters: A developer open-sourced typst-structure within 19 hours — a validation tool that extracts resolved layout structure . Community diagnosis points to Syntactic Frequency Bias: Markdown acts as a "gravitational attractor" .
🟢 UX & Feature Gaps
Issue Details Status
Edit/regeneration limits too strict (#1372, #1373, #1375) Breaks creative workflows No official response
No conversation search Ctrl+F only searches visible screen area Documented
No token counter (#1151, #1050) No visibility into context usage Documented
No cross-conversation reference (#1206) "Remember that chat about my son's karate?" Documented
RTL text direction unsupported (#294, #1378) Arabic, Hebrew not supported No official response
🟢 Community Governance Issues
Issue Details Status
Security reports go unanswered security@deepseek.com has no auto-reply Persistent
No auto-acknowledgment mechanism Community volunteers tracking issues that official team ignores Unaddressed
Stale bot closes active issues #1125 marked stale multiple times Unresolved
💰 Bounty Program – XMR Rewards
To encourage community contributions, bounties in Monero (XMR) are offered for:
Contribution Type Bounty (XMR)
Security fix (any critical issue) 0.5 XMR
Reproducible PoC for any unconfirmed issue 0.1 XMR
API fix with clean PR 0.2 XMR
Validation tool (like typst-structure) 0.1 XMR
Documentation fix 0.05 XMR
How to claim:
Fork the relevant repo
Submit a PR or open an issue with your fix
Tag @qingkong66 or @fkyah3 in the thread
Payment sent within 48 hours of merge/confirmation
Note: Bounties are community-funded, not official DeepSeek compensation.
Summary
Category Unresolved Count
Security & Safety 9 critical
API & Stability 7
Model Behavior 5+
UX & Features 5+
Governance 3
Total ~30+ issues
How You Can Help
Reproduce any issue listed above and add your findings
Submit a PR with a fix (bounties available!)
Add a +1 comment to issues you've encountered — community support helps prioritize
Write validation tools (like typst-structure for other domains)
Share this post with developers interested in open-source AI
Data compiled from GitHub issues and community reports (April–July 2026). Many issues remain without official response.
Top comments (0)