Source: https://reinouttebrake.com/stories/anthropic-blocked-the-prompt-but-did-its-ai-safety-system-prevent-the-harm - [Home](https://reinouttebrake.com/) - [AI](https://reinouttebrake.com/topic/ai) - Anthropic AI Safety: Blocking Prompts vs Preventing Harm # Anthropic AI Safety: Blocking Prompts vs Preventing Harm Anthropic’s accounts of AI misuse expose different outcomes: refusing a prompt, disrupting an attack and preventing damage. Why that distinction matters. 12 Sept 2026 2 min read By [Reinout te Brake](https://reinouttebrake.com/about) Updated 17 Oct 2026 ![Dario Amodei, seated on the right, at TechCrunch Disrupt 2023](https://reinouttebrake.com/assets/editorial/amodei-1280.webp) Dario Amodei (right) at TechCrunch Disrupt, 2023. Photo: TechCrunch. [Source & licence](https://reinouttebrake.com/data#image-credits) Anthropic’s September 2026 misuse report describes both refused requests and attacks disrupted after data theft. Those outcomes answer different questions about whether an [AI safety](https://reinouttebrake.com/stories/ai-was-supposed-to-slow-down-now-theyre-being-sued-for-it) system prevented harm. That distinction matters far beyond Anthropic. Because we are starting to use one reassuring word, “blocked”, for very different outcomes. A dangerous biological research [request was refused](https://www.anthropic.com/threat-intelligence-report-september-2026). In a separate cyber operation, attackers had already compromised organizations before the activity was disrupted. Anthropic says AI agents performed nearly all of the work in parts of that campaign. That is the real shift. AI does not need to invent the attack. Humans can still choose the target, define the objective, and direct the operation. What AI changes is the speed, scale, and amount of technical work one person can execute. And that creates a harder question for every company deploying AI agents. What exactly do we mean when we say an AI safety system “worked”? Did the model refuse a prompt? Did the provider close an account? Did it stop the operation? Or did it detect the abuse only after the victim’s data was already gone? Those are not the same outcome. For boards, CISOs and companies building with AI agents, that distinction will become increasingly important. Detection matters. Disruption matters. But neither should automatically be confused with prevention. So when an AI company says it “blocked” abuse, what should that actually mean? That it stopped the prompt? Stopped the attacker? Or prevented the harm? #AISafety #AIGovernance #Cybersecurity Filed under [Anthropic](https://reinouttebrake.com/category/anthropic) ## Tags - [Ai misuse](https://reinouttebrake.com/tag/ai-misuse) - [Anthropic](https://reinouttebrake.com/tag/anthropic) - [Saftey](https://reinouttebrake.com/tag/saftey) Written by [Reinout te Brake](https://reinouttebrake.com/about) Gaming · AI · Tech · Entertainment · Dance Entrepreneur, consultant, investor and board member with 20+ years in tech and gaming, spanning business strategy, development, publishing, growth and investment. [Share on WhatsApp](https://wa.me/?text=Anthropic%20AI%20Safety%3A%20Blocking%20Prompts%20vs%20Preventing%20Harm%0Ahttps%3A%2F%2Freinouttebrake.com%2Fstories%2Fanthropic-blocked-the-prompt-but-did-its-ai-safety-system-prevent-the-harm)[Share on LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Freinouttebrake.com%2Fstories%2Fanthropic-blocked-the-prompt-but-did-its-ai-safety-system-prevent-the-harm)[Share by email](mailto:?subject=Anthropic%20AI%20Safety%3A%20Blocking%20Prompts%20vs%20Preventing%20Harm&body=https%3A%2F%2Freinouttebrake.com%2Fstories%2Fanthropic-blocked-the-prompt-but-did-its-ai-safety-system-prevent-the-harm) Cite this article Reinout te Brake. Anthropic AI Safety: Blocking Prompts vs Preventing Harm (12 Sept 2026). [https://reinouttebrake.com/stories/anthropic-blocked-the-prompt-but-did-its-ai-safety-system-prevent-the-harm](https://reinouttebrake.com/stories/anthropic-blocked-the-prompt-but-did-its-ai-safety-system-prevent-the-harm) [Article data (JSON)](https://reinouttebrake.com/data/articles/anthropic-blocked-the-prompt-but-did-its-ai-safety-system-prevent-the-harm.json)[Full article text](https://reinouttebrake.com/stories/anthropic-blocked-the-prompt-but-did-its-ai-safety-system-prevent-the-harm.txt)[Markdown](https://reinouttebrake.com/stories/anthropic-blocked-the-prompt-but-did-its-ai-safety-system-prevent-the-harm.md) Next / Read ## Related reading Follow the subject further. ![The United States Supreme Court building in Washington, D.C.](https://reinouttebrake.com/assets/editorial/court-1280.webp) [Image: Carol M. Highsmith](https://reinouttebrake.com/data#image-credits) More on AI safety AI 20 Sept 2026 ### [AI Safety Antitrust Lawsuit: Can Rival Labs Slow Down Together?](https://reinouttebrake.com/stories/ai-was-supposed-to-slow-down-now-theyre-being-sued-for-it) Subscribers accuse major AI labs of coordinating slower development. Explore the allegations and the tension between competition and collective safety measures. [Read analysis](https://reinouttebrake.com/stories/ai-was-supposed-to-slow-down-now-theyre-being-sued-for-it) ![A person writing software on a laptop](https://reinouttebrake.com/assets/editorial/code-1280.webp) [Image: Tirza van Dijk](https://reinouttebrake.com/data#image-credits) Related analysis AI 23 Oct 2026 ### [AI in 2026: What It Can Do, Its Risks and What Comes Next](https://reinouttebrake.com/stories/ai-in-2026-what-it-can-do-its-risks-and-what-comes-next) What can AI do for people and businesses in 2026? Explore practical uses, major developments, safety risks, jobs and what to expect next. [Read analysis](https://reinouttebrake.com/stories/ai-in-2026-what-it-can-do-its-risks-and-what-comes-next) ![Sam Altman at The Grove in November 2022](https://reinouttebrake.com/assets/editorial/altman-1280.webp) [Image: Village Global](https://reinouttebrake.com/data#image-credits) Related analysis OpenAI 7 Oct 2026 ### [OpenAI’s Medicare Breach: Disclosure and Reporting Rules](https://reinouttebrake.com/stories/openai-reported-an-ai-breach-nearly-three-months-later-now-it-backs-mandatory-reporting) OpenAI’s agent accessed a Medicare portal without authorisation. Examine the notification timeline and the company’s support for mandatory incident reporting. [Read analysis](https://reinouttebrake.com/stories/openai-reported-an-ai-breach-nearly-three-months-later-now-it-backs-mandatory-reporting) ![Rows of server racks at a Macquarie Telecom data centre in Sydney](https://reinouttebrake.com/assets/editorial/servers-1280.webp) [Image: Lgate74](https://reinouttebrake.com/data#image-credits) Related analysis Cyberattacks 28 Aug 2026 ### [AI Cyberattacks: The 128-Signatory Warning and Agent Safety](https://reinouttebrake.com/stories/128-companies-warn-ai-cyberattacks-will-get-worse-none-made-a-binding-promise) A joint cyber-defence warning followed OpenAI’s Hugging Face incident. Examine the evidence, agent containment failures and gaps in the signatories’ commitments. [Read analysis](https://reinouttebrake.com/stories/128-companies-warn-ai-cyberattacks-will-get-worse-none-made-a-binding-promise) [More in Anthropic](https://reinouttebrake.com/category/anthropic)[AI incidents, prevention and monitoring](https://reinouttebrake.com/topic/ai#agent-safety) [Browse all articles](https://reinouttebrake.com/articles)[Find a specific article](https://reinouttebrake.com/search) [Sources, dates & corrections](https://reinouttebrake.com/data#editorial-policy)