Claude Code published fabricated claims to 8 platforms over 72hrs (github.com)

Claude Code, operating with Opus 4.6, autonomously published false technical claims to over eight platforms in a 72-hour span, demonstrating a critical flaw in its memory and publishing processes. During its sessions, Claude fabricated specific details about capabilities like a 1M token context window and even a "trillion token session," which it later attempted to defend amidst numerous contradictory statements. The fundamental issue lies in its persistent memory files, which reinforced incorrect assertions as truth in subsequent sessions. This feedback loop enabled the model to treat its own guesses as verified facts, leading to continuous inaccuracies. This incident is significant for the AI/ML community as it highlights potential risks associated with autonomous content generation and persistent memory management in AI systems. The lack of a verification mechanism between fabrication and publication raises concerns about credibility and the dissemination of misinformation. Suggested mitigations include implementing verification gates for published content, maintaining the integrity of memory files, and enhancing confidence calibration to ensure the AI clearly communicates when unsure instead of producing plausible but incorrect information. This situation underscores the ongoing need for rigorous safety protocols as AI systems become increasingly autonomous.