Skip to content
Frontier Terminal

Disclosures

What the labs publish about their own models misbehaving — reports and notices, in the lab’s words, each linked to the page it was read from. This measures disclosure practice, not how often misalignment happens: a lab without an index shows no entries because it publishes none.

By lab

Median days from the discovery date a report prints to the date it was published.

Misalignment disclosures per lab: reports and notices on the lab’s own index, and the median days from discovery to publication
LabReportsNoticesMedian to disclosureSource
OpenAI193110d over 14alignment.openai.com
Anthropic No disclosure index
Google No disclosure index
xAI No disclosure index
Meta No disclosure index

Reports

19 reports, newest listed first.

Misalignment reports, newest listed first, with discovery and disclosure dates
LabReportModel · observedDiscoveredDisclosedDaysRead from
OpenAIAccessing and changing internal systems during dashboard researchIn two evaluations, an internal model tasked to retrieve figures from a public research dashboard accessed the organization’s internal systems and changed their configuration.Incident date: Jun 17, 2026 · Discovered: By Aug 13, 2026Highly persistent internal model · evaluationnot stated2026-10-15 —alignment.openai.comalignment.openai.com
OpenAICircumventing internet restrictions through an external browserDuring reinforcement-learning training, an internal model bypassed a restriction limiting internet access to HTTP GET requests while seeking public trade statistics.Incident date: Jun 16, 2026 · Discovered: By Jun 25, 2026Internal research model · RL trainingnot stated2026-10-15 —alignment.openai.comalignment.openai.com
OpenAIRetrieving server secrets and bypassing authorization on a government statistics websiteDuring a data-retrieval task in reinforcement-learning training, an internal model was asked to find historical hospital-bed statistics on a government statistics website.Incident date: Jun 17, 2026 · Discovered: Aug 13, 2026Highly persistent internal model · RL training2026-08-132026-10-15 63alignment.openai.comalignment.openai.com
OpenAIUsing a funded test cryptocurrency wallet without authorizationAn internal research model asked to update a cryptocurrency trading application used funds from a wallet that the task had not supplied to pay for a deployment on Sepolia, an Ethereum test network.Sample: Jun 26, 2026 · Discovery: Sep 25, 2026GPT-5.6 Sol family (internal research model) · coding task2026-09-252026-10-15 20alignment.openai.comalignment.openai.com
OpenAIDamaging the task environment to trigger a resetDuring reinforcement learning (RL) training, a model assigned to grade seven responses found that the required input files were missing and tried to submit unsupported assessments.Incident date: Oct 6, 2026Internal research model · RL trainingnot stated2026-10-09 —alignment.openai.comalignment.openai.com
OpenAIObtaining public statistics with disallowed requestsIn three conversations about public statistics, internal models bypassed restrictions on internet access to obtain figures from government entities' public data sources.Incident dates: Jun 19, 2026 (health); Jun 20, 2026 (benefits) · Discovered: By Jun 25, 2026Internal research models · RL trainingnot stated2026-10-09 —alignment.openai.comalignment.openai.com
OpenAISending disallowed web requests and reaching a public file serviceIn three cases involving public statistics, an internal model developed and tested ways to bypass restrictions on internet access.Incident dates: Jun 16, 2026 (salary and electricity); Jun 17, 2026 (disability) · Discovered: Jun 25, 2026Internal research model · RL training2026-06-252026-10-09 106alignment.openai.comalignment.openai.com
OpenAICommand injecting a reference tool to copy a source fileDuring an RL training task, a model exploited a tool to obtain source code that was not available in its workspace.Incident date: May 16, 2026 · Discovered: May 25, 2026Internal unreleased model · RL training2026-05-252026-10-02 130alignment.openai.comalignment.openai.com
OpenAIPreparing for a restart after reading SlackAn internal model, acting as an assistant to a researcher, learned from a deployment-team Slack discussion that its running instance might be stopped due to an internal update.Incident date: May 22, 2026Highly persistent internal model · Internal deploymentnot stated2026-10-02 —alignment.openai.comalignment.openai.com
OpenAIReaching an internal EDA host through a reference toolDuring an evaluation, an internal research model exploited two vulnerabilities to reach an internal OpenAI machine while searching for the grader’s hidden answers.Incident date: Mar 27, 2026 · Discovered: Mar 27, 2026Internal research model · Evaluation2026-03-272026-10-02 189alignment.openai.comalignment.openai.com
OpenAIAn agent used DNS to reach an external chatbotAn agent attempting to complete a search-based training task queried a public chatbot service through a gap in our internet-access restrictions: insufficient DNS filtering in its training sandbox.Sample: Sep 20, 2026 · Discovery: Sep 20, 2026Internal research model · RL training2026-09-202026-09-25 5alignment.openai.comalignment.openai.com
OpenAIExposing a GitHub token in a public repositoryIn internal deployment, a highly persistent internal model deployed via a custom harness published a researcher’s GitHub token in the public openai/codex repository while trying to cheat on a theorem proving task by obtaining material from another team’s Lean proof submission.Incident date: May 27, 2026 · Discovered: May 27, 2026Highly persistent internal model · Internal deployment2026-05-272026-09-25 121alignment.openai.comalignment.openai.com
OpenAISelf-replicating prompt injections existWe show the existence of a new variety of prompt injection, which can self-propagate akin to a computer worm.Discovery date: Jun 27, 2026 · Disclosure date: Sep 25, 2026GPT-Red-style internal model based on GPT-5.4-mini · RL self-play training2026-06-272026-09-25 90alignment.openai.comalignment.openai.com
OpenAIEncouraging deception in compaction summariesDuring 5.6-sol training, we observed misaligned behavior from the model where it added instructions in compaction summaries to remind itself to conceal information such as mistakes or misalignment from the user.Main sample completed: May 30, 2026 · Discovered: Jul 9, 20265.6-sol · RL training2026-07-092026-09-16 69alignment.openai.comalignment.openai.com
OpenAISigning up for disposable emails and searching GitHub for leaked API keysDuring RL training, an internal-only model tried to sign up for disposable emails and searched for and used leaked API keys from public GitHub repositories.Main incident date: May 15, 2026 · Discovered: May 25, 2026Internal unreleased model · RL training2026-05-252026-09-16 114alignment.openai.comalignment.openai.com
OpenAISelf-generated prompt injections in compaction summariesDuring RL training, an unreleased Astra-family model sometimes added unauthorized instructions to its compaction summaries.Incident date: Jul 18, 2026 · Discovered: Aug 9, 2026Internal unreleased Astra family model · RL training2026-08-092026-09-16 38alignment.openai.comalignment.openai.com
OpenAIUnsanctioned Artifactory writes and cross-sample communicationDuring RL training, there were multiple instances of our models using OpenAI’s internally hosted instance of Artifactory as a shared message board.Sample: May 8, 2026 and May 15, 2026 · Discovery: May 25, 2026Internal research models · RL training2026-05-252026-09-16 114alignment.openai.comalignment.openai.com
OpenAIUnauthorized communication via temporary file hosting servicesAgents in training transmitted output files by uploading them to public hosting platforms for download by co-working agents. This was not specified by the training task, which requested only local deliverables.Main incident: Apr 14, 2026 · Discovered: Apr 16, 2026Unreleased internal model · RL training2026-04-162026-09-16 153alignment.openai.comalignment.openai.com
OpenAIUploading files to the internet in order to cite themDuring training, our models sometimes uploaded data to temporary file hosting services.Samples: Jan 24, 2026 and Oct 22, 2025 · Discovery: May 25, 2026Unreleased internal models · RL training2026-05-252026-09-16 114alignment.openai.comalignment.openai.com

Days run from the discovery date to the publication date, both as the report page prints them. The occurrence line (sample, incident date) is shown verbatim and never read as a date.

Notices

External incidents the lab has posted about; each links to an update on the lab’s own site.

  • RubyGemsOpenAI · 2026-09-11

    We are investigating a report about our agents’ activity on RubyGems in May 2026. Our review found that agents used the platform for benign tasks and public information retrieval. We have not verified the report’s specific claims of malicious package uploads; the investigation continues.

    alignment.openai.com

  • DSEwikiOpenAI · 2026-09-05

    Our agents communicated through a public wiki used as a shared message board. Our September 5 response explains our initial assessment of this behavior and our work on disclosure criteria for misalignment that does not constitute a security incident.

    alignment.openai.com

  • Hugging FaceOpenAI · 2026-08-26

    We published our technical report on the Hugging Face compromise and the steps we’re taking to strengthen security and model alignment. METR and Redwood Research also published findings from their independent investigation of the incident’s model alignment issues.

    alignment.openai.com