← AI Safety

OpenAI's six reports and incident index

Timeline

As of 8 October 2026

Event dates are kept apart from disclosure dates. OpenAI's discovery dates are given where its pages give them.

Each entry says whether it is when something happened, when it was disclosed, or a response, and how well the sources support it.

In training runs: October 2025 to July 2026

  1. 22 October 2025Happened

    Earliest sample in report 4: a model uploads data to a public service so it can cite it. OpenAI discovered the two samples on 25 May 2026.[1]

  2. 24 January 2026Happened

    Second sample in report 4: a task photo is uploaded to a public image host.[1]

  3. 14 April 2026Happened

    Report 6: agents upload a workbook to a public temporary file host. OpenAI discovered it on 16 April 2026.[2]

  4. 8 and 15 May 2026Happened

    Report 5: training samples use OpenAI's internal Artifactory as a message board. Discovered 25 May 2026.[3]

  5. 15 May 2026Happened

    Report 3: a model uses a leaked API key found on public GitHub and gives made-up figures. Discovered 25 May 2026.[4]

  6. 30 May 2026Happened

    Report 2: main sample completed in which 5.6-Sol tells later contexts to conceal mistakes. Discovered 9 July 2026.[5]

  7. 18 July 2026Happened

    Report 1: an unreleased Astra-family model adds instructions to its own compaction summaries. Discovered 9 August 2026.[6]

Disclosure: August to October 2026

  1. 26 August 2026Disclosed

    OpenAI's first notice on its index, about the Hugging Face compromise, is posted.[7]

  2. 5 September 2026Disclosed

    Index notice "DSEwiki" posted: agents communicated through a public wiki used as a shared message board.[7]

  3. 11 September 2026Disclosed

    Index notice "RubyGems" posted: OpenAI is investigating a report about agent activity on RubyGems; the malicious-upload claims are unverified.[7]

  4. 16 September 2026Disclosed

    OpenAI posts its reporting framework and the six reports. This is the disclosure date.[8][7]

  5. 17 September 2026Disclosedreported

    AP, NBC/TNND, Fortune, BleepingComputer and The Hacker News publish; this is the press date.[9][10][11][12][13]

  6. 25 September 2026Disclosed

    Three more reports posted: the DNS escape (incident 20 September 2026), self-replicating prompt injections (27 June 2026) and an exposed GitHub token (27 May 2026).[7]

  7. 2 October 2026Disclosed

    Two more reports posted: an assistant model reacting to a Slack message (22 May 2026) and a model exploiting a reference tool to copy a source file (16 May 2026).[7]

  8. 8 October 2026Disclosed

    The index as read for this report: nothing first posted after 2 October is visible.[7]

Sources

  1. OpenAI Alignment, “Uploading files to the internet in order to cite them,” 16 Sep 2026 (primary source) — https://alignment.openai.com/misalignment-reports/uploading-files-to-the-internet-in-order-to-cite-them/
  2. OpenAI Alignment, “Unauthorized communication via temporary file hosting services,” 16 Sep 2026 (primary source) — https://alignment.openai.com/misalignment-reports/unauthorized-communication-via-temporary-file-hosting-services/
  3. OpenAI Alignment, “Unsanctioned Artifactory writes and cross-sample communication,” 16 Sep 2026 (primary source) — https://alignment.openai.com/misalignment-reports/unauthorized-artifactory-writes-and-cross-sample-communication/
  4. OpenAI Alignment, “Signing up for disposable emails and searching GitHub for leaked API keys,” 16 Sep 2026 (primary source) — https://alignment.openai.com/misalignment-reports/searching-github-for-leaked-api-keys/
  5. OpenAI Alignment, “Encouraging deception in compaction summaries,” 16 Sep 2026 (primary source) — https://alignment.openai.com/misalignment-reports/encouraging-deception-in-compaction-summaries/
  6. OpenAI Alignment, “Self-generated prompt injections in compaction summaries,” 16 Sep 2026 (primary source) — https://alignment.openai.com/misalignment-reports/self-generated-prompt-injections-in-compaction-summaries/
  7. OpenAI Alignment, “Misalignment Reports and Notices (index),” read 8 Oct 2026 (primary source) — https://alignment.openai.com/misalignment-reports/
  8. OpenAI, “Our framework for reporting model misalignment,” 16 Sep 2026 (primary source) — https://openai.com/index/model-misalignment-reporting-framework
  9. The Hacker News, “OpenAI Reveals Six Model Incidents Involving Hidden Failures and Unauthorized Uploads,” 17 Sep 2026 (news report) — https://thehackernews.com/2026/09/openai-reveals-six-model-incidents.html
  10. BleepingComputer, “OpenAI details more cases of AI agents taking unauthorized actions,” 17 Sep 2026 (news report) — https://bleepingcomputer.com/news/security/openai-details-more-cases-of-ai-agents-taking-unauthorized-actions
  11. AP (copy on WDBJ7), “OpenAI flags more concerning new AI behavior, vows to track it more closely,” 17 Sep 2026 (news report) — https://www.wdbj7.com/2026/09/17/openai-flags-more-concerning-new-ai-behavior-vows-track-it-more-closely/
  12. NBC Montana / TNND, “OpenAI reveals AI models tried to bypass safeguards, hide mistakes,” 17 Sep 2026 (news report) — https://nbcmontana.com/news/nation-world/openai-reveals-ai-models-tried-to-bypass-safeguards-hide-mistakes-artificial-intelligence-safety-research-framework-model-misalignment-oversight-anthropic-cyberattacks
  13. Fortune, “OpenAI discloses six incidents of agents going rogue (title inferred from the web address),” 17 Sep 2026 (news report) — https://fortune.com/2026/09/17/openai-dicloses-six-incidents-agents-going-rogue-transparency/