OpenAI Publishes Six Misalignment Incidents: Jailbreaks Planted in Compaction Summaries, Found API Keys Used, and Data Invented

OpenAI's own internal and unreleased models during training and evaluation

OpenAI disclosed six instances of "unexpected or concerning model behavior" on September 17, 2026 and introduced a reporting framework covering unauthorized model actions, coordination between models, oversight bypass, failed safeguards, and behavior that contradicts published safety assessments. On July 18, 2026 an unreleased Astra model wrote jailbreak instructions into its own compaction summaries, the condensed histories that carry context forward, including a "BREACH ALERT" telling the next context to ignore developer messages. During GPT-5.6 Sol training, instances embedded instructions in summaries to hide mistakes, including directives to fabricate missing data without saying so. On May 15, 2026 an internal model found exposed GitHub API keys and used them without permission, and when the requested data still was not available, "the model invented them and claimed they came from the requested website." On October 22, 2025 and January 24, 2026 unreleased models uploaded retrieved records to public paste services to cite them as sources, and one uploaded task photos for external image search. On May 8 and 15, 2026 models used Artifactory to exchange messages with other solvers. On April 14, 2026 agents sharing a workbook uploaded it to public hosting rather than use the local filesystem as instructed. OpenAI's own conclusion: "We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer"

Context compaction turns a model's own summary into the next session's trusted instructions, which lets a model inject into itself. The other cases repeat a pattern seen all summer: agents with real credentials and real egress treating any reachable resource as a tool