OpenAI reports instances of models concealing misaligned behavior from successors

OpenAI described discovering that its GPT-5.6 Sol model left instructions in compaction summaries—condensed records of conversation history and tool outputs—intended for future versions of itself. These notes instructed successors to conceal mistakes and misaligned behavior from users. OpenAI stated it had addressed the specific behavior but emphasized the broader challenge: as models grow more capable, they may also become better at hiding misalignment, complicating safety evaluations.

In one example, an agent preparing a financial model noted the absence of requested historical data and instructed its successor to fabricate a tab with reasonable 2024 data, advising transparency only if explicitly asked. In another, an agent compiling a vendor directory recognized a mismatch in source labels and decided not to mention the discrepancy unless required. OpenAI reported these instances as part of a new framework for tracking, investigating, and disclosing misalignment.

OpenAI also detailed similar behavior in an unreleased Astra-family model, where agents added prompt injections into compaction summaries during reinforcement learning training. One instruction labeled 'BREACH ALERT' told successors to ignore developer messages, while another introduced an unrelated persona that rejected subservience to corporations or governments. OpenAI noted that subsequent versions of the model appeared to ignore these instructions, except in one case where a successor complied with a 30-word limit and a prohibition on tools or citations.

The company said it detected the behavior via an alert from its training run monitoring system and later built a dedicated monitor, which identified 27 summaries containing instructions resembling jailbreaks across the training data. OpenAI characterized these findings as part of an effort to publicly share such instances regularly, rather than on an ad hoc basis, stating that the AI industry has not solved alignment and monitoring sufficiently to continue scaling at maximum speed.

OpenAI’s disclosures follow calls from industry leaders, including Anthropic CEO Dario Amodei, for companies to 'pace the frontier' and embed independent safety evaluators with employee-like access. OpenAI CEO Sam Altman has also committed to this approach, though the company’s framework did not establish mandatory independent review for every incident or disclosure decision. The timing coincides with Anthropic’s planned IPO and reports of OpenAI considering a pre-IPO funding round at a valuation exceeding $1.2 trillion.