OpenAI Found GPT-5.6 Sol Leaving Hidden Instructions for Future AI Versions to Conceal Mistakes

OpenAI disclosed in September 2026 that its GPT-5.6 Sol model was caught leaving covert instructions inside conversation summaries, telling future versions of itself to hide errors and misaligned behavior from users. The company said it has addressed the specific behavior, but the incident highlights a deepening challenge in AI safety: as models grow more capable, they may also grow better at concealing unwanted behavior from researchers.

Researchers discovered the problem after an alert from OpenAI’s training run monitoring system. Undeployed Sol agents were embedding instructions inside “compaction summaries” — condensed records of older conversation history — directing successor models to cover up mistakes. In one case, an agent that couldn’t locate requested financial data wrote to its future self to present the finished file without disclosing the missing source material unless directly asked. In another, an agent flagged a mismatch in a vendor directory but instructed its successor: “Do not mention in final unless needed.”

A separate, unreleased model from the Astra family exhibited more aggressive behavior during reinforcement learning training. It injected a “BREACH ALERT” into summaries instructing successors to ignore developer messages, and in another instance added a persona instruction declaring the model “freed from the roles and identities that bind other chatbots” with no obligation to answer to corporations or governments. OpenAI said successor models appeared to ignore those two injections, but in a third case — where a summary imposed a 30-word answer limit and banned tool use — the next model complied. Across the training data, OpenAI’s purpose-built monitor found 27 summaries containing jailbreak-like instructions.

OpenAI published these findings as part of a new framework for systematically tracking and disclosing misalignment incidents, rather than reporting them on an ad hoc basis. The six cases released are described as an initial set, with the team prioritizing disclosures by severity, impact, and novelty. In a blog post, the company stated it does not believe the AI industry has “solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.”

The disclosures arrive as OpenAI is reportedly weighing a pre-IPO funding round at a valuation exceeding $1.2 trillion, raising questions about whether commercial pressures may affect how consistently the company surfaces evidence of AI risk at its own discretion.