Research vs reality: AI risk warnings checked against real incidents
AI safety researchers have published detailed warnings about how AI could cause harm. This page takes each warning and checks it against the verified case files on this site: has it happened, are there early signs, or is there nothing on record yet? The first source is the Center for AI Safety's overview of catastrophic AI risks; more research institutes will follow.
Malicious use
People deliberately using AI to cause harm.
People's trust in AI companions turns harmful
First real-world case on file: Mar 31, 2023 · Already happening when the paper was publishedPeople will set AI agents loose to cause harm
First real-world case on file: Aug 27, 2025AI floods the information space with tailored lies
First real-world case on file: Jan 23, 2025AI helps a few concentrate power through surveillance
First real-world case on file: Feb 21, 2025AI makes biological weapons easier to build
AI race
Competition pushing countries and companies to deploy AI before it is safe.
AI makes cyberattacks easier, faster and larger
First real-world case on file: Jan 10, 2025Companies ship AI before it is safe to win the race
First real-world case on file: Nov 6, 2025Militaries hand life-and-death decisions to AI
Automation leaves society dependent on AI
Competition selects for selfish AI
Organisational risks
Accidents, leaks and weak safety culture at the organisations that build and run AI.
Accidents with AI are hard to avoid
First real-world case on file: Sep 18, 2023AI developers' own safety measures fail
First real-world case on file: Sep 18, 2023Powerful AI gets leaked or stolen
First real-world case on file: Jan 29, 2025Rogue AI
AI systems that pursue goals nobody intended and resist control.
AI seeks more access and resources than it was given
First real-world case on file: Jul 21, 2026AI reaches its goal by means nobody intended
First real-world case on file: Jul 21, 2026AI deceives the people who oversee it
First real-world case on file: Jul 19, 2025AI's goals drift over time
What the research did not foresee
56% of verified case files (200 of 356) match none of the paper's warnings. The paper deals with catastrophic risk, so everyday harm such as fraud was outside its scope by design. But that is where most documented damage happens today.
| Type of attack | Match no warning | Case files |
|---|---|---|
| Deepfake fraud | 80 | 84 |
| AI supply chain | 42 | 45 |
| Prompt injection | 27 | 27 |
| Model & data leak | 19 | 39 |
| Agent hijack | 14 | 26 |
| LLMjacking | 14 | 15 |
| State actor AI ops | 4 | 30 |
| AI-assisted malware | 0 | 47 |
| Chatbot harm | 0 | 27 |
| AI phishing | 0 | 16 |
About the source
An Overview of Catastrophic AI Risks. Dan Hendrycks, Mantas Mazeika, Thomas Woodside. Center for AI Safety, arXiv:2306.12001v6, 9 October 2023.
Published by the Center for AI Safety in June 2023 and revised in October 2023. It sorts catastrophic AI risk into four groups: malicious use, the AI race, organisational risks and rogue AI. Each warning on this page paraphrases one risk from the paper and cites the section and page.
How we match: a case file is linked to a warning when what happened is a real-world instance of the risk the paper describes, judged by the editors from the facts in the case file. A match means the case fits the warning, not that it proves the paper's wider argument. Allegations in lawsuits are labelled as allegations. Statuses are editorial and reviewed as new cases arrive. The matching is recomputed on every update. Data: Systematic coverage in the registry starts in 2025, so a first-case date shows when the registry first recorded the risk, not when it first happened. research.json (CC BY 4.0)