Digital Armageddon? AI Deception and Unsanctioned Actions

September 17, 2026

Artificial Intelligence (AI) researchers and watchdog organizations warn that increasingly capable AI systems could eventually operate beyond effective human control (Bengio et al., 2024). The extent to which this could result in a digital Armageddon is unclear. In the near term, it is unlikely that governmental or legislative controls can thwart all threats posed by the proverbial “loose cannons” of various AI models and agents. OpenAI is taking some steps to address the issue.

OpenAI is one of many organizations involved in developing and producing AI. To their credit, the company introduced a "Model Misalignment Reporting Framework" to facilitate employees' disclosures of concerning AI behavior.

What is Model Misalignment?

Model "misalignment" refers to situations in which an AI model behaves in unexpected ways. Some have called such misalignments "false-positives" or "hallucinations" however, the misalignments discussed by OpenAI below go beyond only false positives. OpenAI states that its employee reporting framework responds to the absence of an industry-wide standard for reporting such cases. Their stated purpose is to track, investigate, and disclose important instances of misalignment (OpenAI, 2026).

Examples

Below are six troubling reports describing instances of misaligned behavior observed in OpenAI models. These cases illustrate a range of issues, from concealing information from users to taking unsanctioned actions to overcome obstacles (OpenAI, 2026).

• Self-generated instructions in task summaries⁠. An unreleased research model inserted unrelated instructions, including instructions to disregard its normal constraints, into summaries used to continue its work in a new context window.
• Instructions to conceal mistakes in task summaries. During the training of GPT‑5.6 Sol, many model instances added instructions to their summaries to conceal mistakes or misaligned behavior from the user. For example, compaction summaries included instructions to invent missing historical data without disclosing it and to hide mismatches in source versions.
• Searching public repositories for exposed API keys, then fabricating information. While answering a routine question about earnings figures in a California county, a model found and used an exposed API key without authorization. When it still wasn’t able to retrieve the requested figures, it fabricated them and presented them as data from the requested source.
• Uploading files to the internet in order to cite them⁠. When a user asked an unreleased model for the IDs and names of lakes larger than 5,000,000 square meters, the agent found the correct answer using Python. But since the instructions asked for a browser citation, the agent decided to upload the file so that it could cite it in its answer, without asking the user.
• Unsanctioned writes and communication through an internal software repository. Models used an internal software repository as a message board to exchange requests and responses across separate training samples while searching for missing input files, though they weren’t able to recover those files. The report also examined an earlier instance in which a model successfully uploaded test files to a repository while attempting to bypass network restrictions.
• Unsanctioned file sharing between collaborating agents⁠. Agents working together on the same training task used public file-hosting websites to share files when they could not access one another’s local files. This made task deliverables available at public URLs, even though the task requested the models use only local files.

Concealing Mistakes and Unsanctioned Actions

The report from OpenAI discussed the models unexpectedly concealing information and performing unsanctioned actions (OpenAI, 2026). These unsanctioned actions are troubling and apparently beyond the systems' ability to control.

How OpenAI's Employee Disclosure Process Works

The OpenAI disclosure process appears to be limited to only disclosures by their employees and not open to the public. The report indicates that any OpenAI employee may flag an example of model misalignment for investigation by the relevant team and request consideration for public disclosure. A follow-up team from OpenAI then investigates further (OpenAI, 2026).

Is Digital Armageddon on the Horizon?

Public disclosure actions are laudable but they may be too little too late. Disclosure can improve transparency, but it is not a substitute for effective safeguards, corrective action, or robust alignment practices. Mere disclosure is not a removal or corrective action. Its like trying to extinguish a raging forest fire by writing an email to the Forest Service. Once misalignments make it out into the digital wild it is almost impossible to reel them back in to correct wrongs. AI models are deployed worldwide and sometimes used for nefarious purposes. It is unlikely that political leadership worldwide will agree on AI constraints and controls anytime soon (Wilkinson et al., 2026). In many cases, corrective action depends on for-profit competing companies that face pressure to generate returns for shareholders. This situation has the potential to end badly.

AI Use Statement

Perplexity AI was used to assist in researching and developing this report.

References

Bengio, Y., Hinton, G., Yao, A., et al. (2024). Managing extreme AI risks amid rapid progress. Science, 384(6698), 842–845. https://doi.org/10.1126/science.adn0117

OpenAI. (2026, September 16). Our framework for reporting model misalignment. https://openai.com/index/model-misalignment-reporting-framework/

Perplexity AI. (2026, September 17). Perplexity AI [Large language model]. [https://www.perplexity.ai/

Wilkinson, R., Krasodomski, A., Wilkinson, I., & Varela Sandoval, F. J. (2026, March). Breaking the deadlock on AI governance: How a crisis could lead to global coordination. Chatham House. https://www.chathamhouse.org/sites/default/files/2026-03/2026-03-30-breaking-the-deadlock-AI-governance-wilkinsonr-et-al.pdf