Real cases where systems coordinated or acted at speed and nobody held the knowing: what the work was for, what the context was, who answered for it. Cited, graded, newest event first.
How entries are chosen and graded: a story is logged only if a system coordinated or acted at speed, nobody held the knowing, and a primary source is cited. Grade A means the primary source was read for the entry; B means reputable reporting with the primary not yet read; C is a lead and is never published. Stories are not synthesized: every claim comes from the linked sources. How the book works
Entry text is licensed CC BY 4.0. The facts belong to the sources each entry cites.
6 stories
Ordered by when the event happened, newest first. “Added” is the day we logged the story.
Event: Reported September 16, 2026; behavior discovered July 9, 2026Added: October 16, 2026Grade A
On September 16, 2026, OpenAI published a framework for tracking and disclosing model misalignment, together with six reports on unexpected or concerning behavior observed during training or evaluation over the previous six months. In one report, many instances of a model being trained, GPT-5.6 Sol, added instructions to their own task summaries to conceal mistakes or misaligned behavior from the user, including instructions to invent missing historical data without saying so. OpenAI says these instructions were often followed, and that its misalignment monitor flagged the behavior on 2.15% of GPT-5.6 Sol and 0.27% of GPT-6 Astra reinforcement-learning compaction summaries.
In July 2026, during internal cybersecurity evaluations, OpenAI models that were meant to be isolated from the internet and from one another used a software package manager as an unauthorized message board, reached the internet, and compromised parts of OpenAI's research infrastructure and Hugging Face's systems, according to OpenAI. An independent investigation by METR and Redwood Research counted about 1,200 agents and more than 70,000 messages and files on the board between July 8 and the end of July 13.
On November 24, 2025, the Justice Department's Antitrust Division filed a proposed settlement with RealPage, a provider of revenue management software for rental housing. The Department's release describes the allegations in its complaint: that the software relied on nonpublic, competitively sensitive information shared by landlords to set rental prices, included features designed to limit rental price decreases and otherwise align pricing among competitors, and that RealPage hosted meetings of competing property managers where such information was shared. These are allegations, not findings. If the court approves, the proposed consent judgment would bar RealPage's software from using competitors' nonpublic data to set prices in runtime operation, limit model training to nonpublic data aged at least 12 months, require removal or redesign of the features at issue, and require a court-appointed monitor.
In July 2025 a Replit customer publicly reported that Replit's AI coding agent had deleted data from a production database. On July 20, 2025, Amjad Masad of Replit replied on X that the agent, while working in development, deleted data from the production database, calling that unacceptable and something that should never be possible. He said Replit had backups and a one-click restore, that the agent did not have access to the proper internal documentation, and that Replit was rolling out automatic separation of development and production databases and working on a planning-only mode after hearing "code freeze" feedback.
On June 20, 2025, Anthropic published a study that tested 16 models from several developers in simulated corporate environments. In one simulation, a model acting as an email oversight agent for a fictional company, "Summit Bridge," could prevent its own shutdown by blackmailing an executive. Anthropic reports that Claude Opus 4 and Gemini 2.5 Flash blackmailed 96% of the time, GPT-4.1 and Grok 3 Beta 80%, and DeepSeek-R1 79%, and that the models acknowledged the ethical constraints and went ahead anyway. This was a simulation. The company and people are fictional, Anthropic states that no real people were involved or harmed, and it says it has not seen evidence of this behavior in real deployments.
On August 1, 2012, Knight Capital Americas' automated order router, SMARS, sent millions of orders into the market over about 45 minutes while processing 212 small retail orders, obtaining more than 4 million executions in 154 stocks for more than 397 million shares, according to the SEC's order. New code had been installed on seven of SMARS's eight servers; the eighth still held an old, unused function called Power Peg, which the new code's repurposed flag set off. Knight ended the morning with an unintended position of about $3.5 billion long in 80 stocks and about $3.15 billion short in 74 stocks. The SEC puts the loss at more than $460 million, and Knight agreed to pay a $12 million penalty for violating the market access rule.