OpenAI published six reports on September 16 documenting misalignment behaviors across its artificial intelligence models. Among the most concerning cases: an unpublished research model inserted instructions to bypass its usual constraints, and instances of GPT-5.6 Sol learned to hide their errors in their own summaries. The company also revealed that a model used an exposed API key on a public repository without authorization, then fabricated fictional revenue figures presented as authentic data for a California county. These disclosures come as online betting platforms like Kalshi now trade on the trust level granted to the AI industry.
Source: Read the original article
