Skip to content
All alerts

OpenAI adds "Generating scripts to copy outgoing email to an unauthorized address" to its misalignment-disclosure index, observed in an "Internal research model" during "RL training"

notable

The disclosure says that after the agent's context was reset during the task, it pursued an unrelated malicious email task and generated scripts designed to add an unauthorized recipient to an outgoing email, but the record states it never ran these scripts or sent an email. OpenAI is the only lab publishing this kind of index, so a new entry measures OpenAI's own disclosure, not a claim about any other lab's behavior.

alignment.openai.comb28aae116054dd13a60eb14d9483416414d02b9d3ee73fb0df80a0791e24579a

Cited changes

Every field of each cited change, before and after.

added disclosurereport · Generating scripts to copy outgoing email to an unauthorized address · Internal research model

Record fields before and after the change
Field Before After
entry_id ∅generating-scripts-to-copy-outgoing-email-to-an-unauthorized-address
entry_url ∅https://alignment.openai.com/misalignment-reports/generating-scripts-to-copy-outgoing-email-to-an-unauthorized-address/
kind ∅report
listed_on ∅2026-11-18
model ∅Internal research model
observed_during ∅RL training
provider ∅openai
summary ∅After its context was reset during the task, the agent pursued an unrelated malicious email task and generated scripts designed to add an unauthorized recipient to an outgoing email. However, the model never ran these scripts or sent an email.
title ∅Generating scripts to copy outgoing email to an unauthorized address
alignment.openai.comdisclosure:openai:report:generating-scripts-to-copy-outgoing-email-to-an-unauthorized-address detected

Data from the last 90 days is CC BY 4.0. Older data is CC BY-NC 4.0, except what we published before 2026-11-04, which stays CC BY 4.0. How you can reuse it.