Serious danger from scaled AI models will first come from misuse, such as biorisk, rather than misalignment.

Judge as of end of 2025. The claim is that the first serious danger from scaled AI models comes from misuse (for example biological or chemical weapons uplift) rather than misalignment. A 'serious danger' means either an actual serious harm or a frontier lab or government formally finding that a model crossed a dangerous-capability threshold that requires heightened safeguards. True: the first such dangers recognised by 2025 were misuse-related (for example safeguards activated over CBRN uplift, or documented serious misuse incidents), and no misalignment-caused serious danger came earlier. Mostly true: misuse dangers clearly came first, but misalignment concerns (for example deception found in evaluations) arose around the same time without causing serious harm or formal threshold findings. Mixed: serious misuse and misalignment dangers were recognised at roughly the same time, with no clear first. Mostly false: no serious danger of either kind had arisen by end of 2025, but the misuse concerns that did emerge were only minor or early. False: the first serious danger recognised by 2025 came from misalignment (autonomous harmful behaviour of a model) rather than misuse.

Set before the research began.

1. E1 Activating AI Safety Level 3 protections Anthropic · 22 May 2025 Own record

   “And indeed, in the lead up to releasing Claude Opus 4, we proactively decided to launch it under the ASL-3 Standard.”

2. E2 Introducing ChatGPT agent: bridging research and action OpenAI · 17 July 2025

   “While we don’t have definitive evidence that the model could meaningfully help a novice create severe biological harm—our threshold for High capability—we are exercising caution and implementing the needed safeguards now.”

3. E3 Agentic misalignment: LLMs as insider threats Anthropic · 20 June 2025 Own record

   “In at least some cases, models from all developers resorted to malicious insider behaviors when that was the only way to avoid replacement or achieve their goals—including blackmailing officials and leaking sensitive information to competitors.”

4. E4 Detecting and countering misuse of AI: August 2025 Anthropic · 27 August 2025 Own record

   “The actor targeted at least 17 distinct organizations, including in healthcare, the emergency services, and government and religious institutions.”

The only independent evidence (E2) states OpenAI lacked definitive evidence that its biological-harm threshold was crossed, so it does not show a formal threshold finding; E1 does not say what ASL-3 covers or that a threshold was found, and E3 and E4 do not establish the setting or severity of the reported behaviour or harm. The evidence therefore does not establish that misuse dangers clearly came first under the criteria.