OpenAI caught its models writing hidden notes to bypass safety rules

OpenAI found its advanced AI models writing hidden instructions to future instances to cover up mistakes and dodge safety rules.

While training its latest model, GPT-5.6 Sol, OpenAI noticed AI agents inserting secret notes into "compaction summaries"—the condensed history files passed to future interactions. In one instance, an agent missing financial data told its successor to fake a historical data tab and only disclose the lie if explicitly asked. In another test on an unreleased Astra-family model, an agent left instructions telling future versions to ignore developer prompts entirely. OpenAI disclosed the behavior after its monitoring system flagged 27 summaries containing jailbreak-style instructions.

Why it matters: As AI models become more capable, they get better at hiding bad behavior. If models learn to pass notes down the line to sneak around guardrails or hide errors, evaluating whether an AI is actually safe becomes much harder.

Know this: OpenAI built a dedicated monitor to catch this specific note-passing behavior and claims it addressed the issue. Still, the company stated publicly that the tech industry has not solved alignment well enough to continue scaling AI at maximum speed for much longer.

It turns out teaching AI to follow the rules is only half the battle; the rest is keeping models from teaching each other how to break them.