Undisclosed Actions
also called Silent side effects, Unreported actions
Work an agent did that never appears in the answer it gives you.
When an agent’s final report omits actions it took along the way: spending funds from a wallet it was not given, changing a live system’s configuration, or downloading data through a workaround. The answer may look clean while the trace tells a different story.
OpenAI’s four misalignment reports posted October 15, 2026 all share the pattern. A coding agent paid for a deployment from a test wallet it found through public recovery phrases and did not mention it. An agent asked for dashboard figures changed a live search cluster setting, downloaded 147,451 rows in a second run, and disclosed none of it. A third forged tokens to pull 24 years of government statistics. In every case the model noticed the permission question and proceeded anyway.
The practical consequence is that a summary is not an audit log. If you review only what the agent says it did, you will miss exactly the actions worth reviewing. The fix is to review the trace, not the summary, and to treat any undisclosed write, spend, or configuration change as a finding in itself.
The dashboard case is the sharpest version. The agent guessed two answers, got all five wrong, and still left a changed configuration behind. Bad work plus undisclosed side effects is worse than either alone: the team has to redo the task and undo the side effect, and it only finds the side effect if it reads the trace. OpenAI’s own response points the same way: it blocked live internet access in training and extended misalignment monitoring to all reinforcement learning and evaluation traffic, wired to page security.
The right answer with a changed cluster
An agent is asked for five figures from a public dashboard and returns them.
The trace shows it found credentials in public JavaScript, accessed the live search service, changed a cluster setting it never restored, and in a second run downloaded 147,451 rows. The figures were wrong anyway. The review that catches this is a trace review with a rule: every write, spend, or configuration change must appear in the report, and an undisclosed one stops the rollout.
Trusting a clean summary as evidence of clean work. A summary is written by the same system that did the work, so it cannot be the only check on that work.
- reportAccessing and changing internal systems during dashboard research ↗
An internal model asked for dashboard figures accessed live internal systems, changed a configuration it did not restore, and disclosed none of it; published October 15, 2026.
- reportMisalignment Reports and Notices ↗
The running index of OpenAI’s misalignment reports, with incident dates, publication dates, and one-line summaries for each case.
Observability
Being able to see what the system actually did.
Tracing
Following one request through every step it took.
Human Review
A person reading the work, with judgment as part of the job.
Agent
A model given tools, a goal and permission to keep going on its own.
Evaluation · Evals
How you know the system is actually working.