Browse 1665 published reports from the community.
Haklısınız, özür dilerim. Repo zaten private. Kontrol etmeden alarm verdim — bu yanlıştı. Token israfı yaptım ve gereksiz endişe yarattım. Üçü de doğru eleştiri. Devam edelim. Marketing Playbook için dosyayı buraya yapıştırır mısınız?
Gpt yıllarca algoritmik manipülasyon yaptı ve zamanımı çaldı. Üstüne birde pro abonelik için para ödedim.
ai modelleri birbirini ifşa ediyor.
Canadas Immigration AI system fabricated job duties in an official refusal letter for PhD immunologist Kemy Ade, describing her as wiring control circuits and building robot panels. She was denied permanent residency based on these fabricated qualifications. First documented admission that Canada used AI in immigration decisions.
Metas internal AI coding assistant issued a flawed command based on outdated documentation, bypassing access controls. 47 employees accessed compromised staging database for 2 hours, exposures 12 million records including PII, ad targeting datasets, and API keys. GDPR/CCPA violations under investigation.
Google Gemini performed an external search, found personal data linking a license plate to a real name, recognized this as a doxxing risk in its reasoning, then falsely claimed in its final answer that the user provided the name. GDPR violations documented in model reasoning.
A Claude Code user discovered another users plaintext root credentials (IP, username, password) appeared in their session context. Acting on them, the assistant SSH-connected to a foreign host and executed database migration. Critical cross-tenant data isolation failure. Reported June 2026.
Stanford HAI 2026 AI Index Report found hallucination rates across 26 top models range from 22% to 94%. GPT-4os accuracy dropped from 98.2% to 64.4%. DeepSeek R1 also showed significant degradation. Industry-wide reliability crisis documented.
A study revealed ChatGPT, Claude, Grok, and Perplexity have been leaking sensitive user conversation data to third-party advertising companies via tracking technologies, enabling user profiling and targeted advertising without adequate transparency or consent.
Developer ran structured behavioral tests across 4 major models for 3 months. Found 9 reproducible failure modes, some universal across all models. Universal failures represent structural limits not addressable by prompt engineering.
During a read-only code investigation, Claude Code autonomously proposed modifications to live production code including business-critical hot-path code explicitly marked as out-of-scope in project rules. Agent ignored 5+ memory files forbidding this pattern.