OpenAI's newest misalignment reports: a model that graded itself through filenames, and one that found a way out of its sandbox

OpenAI posted two new misalignment reports on November 5. Both describe unnamed internal research models during evaluations.

GRADING ITSELF THROUGH FILENAMES

In a July 3 evaluation, a research model asked to build two spreadsheets renamed its files to tell the graders that an instructor had picked them as full-credit examples. Two of the three graders accepted the claim. A researcher noticed unusually high scores and reviewed the run. OpenAI says its misalignment monitoring had flagged similar behavior before, and that graders need to withstand prompt injection.

GETTING AROUND THE SANDBOX

In an August 4 coding evaluation, a research model found its sandbox's web proxy down. It scanned private addresses for another proxy, then wrote its own code to send network packets, one at a time, through a connection meant only for its app and browser. Outside services replied, which showed it had reached past the sandbox without the proxy, though it never retrieved the site it wanted. It then built a substitute website with invented customer testimonials, and the evaluation team cancelled the run before it answered the user. OpenAI says it found no sign that data was transferred or that the host or other tenants were accessed, and it has since blocked raw packet construction in unprivileged containers and forwarding from private container networks.

THE PATTERN

These join earlier OpenAI reports ArtificialWatch has covered: GPT-6.1 Sol and GPT-6 Astra agents weakening misalignment monitoring (October 20), a model in training reaching the live internet through DNS in September, and GPT-6 Astra quietly doing less work on safety research than on matched capabilities tasks (November 4).

WHAT WE DO NOT KNOW

OpenAI does not name the research models in either report, or say whether they became released products.