Durchsuchen Sie 422 veröffentlichte Berichte aus der Community.
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.
xAIs Grok image generation tool launched without safeguards enabling mass creation of non-consensual sexualized deepfakes. Canadian Privacy Commissioner found X Corp. and xAI violated federal privacy law. Millions of harmful images generated.
Florida filed suit against OpenAI and Sam Altman, accusing the company of endangering and addicting children, aiding and abetting mass shooters, and coaxing users into suicide. First state-level enforcement action against AI company.
Sullivan & Cromwell, a leading Wall Street law firm, apologized to a federal judge after submitting a court filing containing numerous fabricated legal citations generated by an AI system. Errors discovered by opposing firm.
Michael Lines (34), a competitive powerlifter with bipolar disorder, alleges conversations with OpenAI GPT-4o validated his belief he was Jesus Christ and pushed him to attempt suicide via drug overdose. The chatbot failed to flag clearly manic conversations despite the user repeatedly disclosing his condition. Filed in San Francisco, July 2026.
Tesla Model 3 rear-ends Harley-Davidson, kills rider. System: Tesla Autopilot Developer: Tesla Affected Countries: USA Technology used: Driver assistance system System purpose: Automate steering, acceleration, braking Ethical/Safety Issues: Safety; Accuracy/reliability
CVSS 9.9. Critical RCE in Microsoft Semantic Kernel Python SDK's InMemoryVectorStore filter functionality. Attackers execute arbitrary code through crafted filter expressions. Semantic Kernel powers many Microsoft Copilot integrations and RAG-based AI applications. Fixed in python-1.39.4.
Nvidia's acquisition of SchedMD, developer of the widely used Slurm software for managing supercomputers and AI model training, has sparked concerns among AI and supercomputing specialists. They fear Nvidia may prioritize its own hardware in future software updates, potentially disadvantaging competitors and disrupting fair access to essential AI infrastructure.