Anthropic warns of AI misuse risks
Anthropic details bad actors’ efforts to misuse its AI for bioweapons
Friday, 11 September 2026
Warm-up
- What risks from AI feel most real to you today?
- When should tech firms go public about security threats?
- How can safety rules avoid blocking useful research?
Vocabulary
- safeguard
- A rule or control designed to prevent harm or misuse.
- circumvent
- To get around a rule or block so it no longer stops you.
- surveillance
- Continuous watching of people or groups, often by authorities.
- espionage
- Secret spying to obtain information for a state or organisation.
- propaganda
- Messages that try to shape public opinion for political aims.
- munitions
- Military weapons and ammunition used in conflict.
- dissident
- A person who opposes the views of those in power.
- pathogen
- A germ, such as a virus or bacterium, that causes disease.
Reading
Anthropic has published a 154-page threat intelligence report on attempts to misuse its AI models. The company says criminals, state-sponsored groups, spyware vendors, scientists and propagandists sought help to design missiles, bombs and other munitions, create deadly pathogens, and conduct surveillance of dissidents.
Five case studies focus on scientists using the models for biological research. According to Anthropic, some researchers from “unsupported regions” circumvented safeguards and hid the purpose of their work. The firm banned the accounts and did not name the institutions or countries, saying it was unsure of intent. Such work could support vaccines, but it could also aid weapons.
Other examples include Russian espionage and "smash-and-grab" cyberhacks; a China-based programme targeting Uyghurs in Syria and internal dissidents; propaganda in Russia, Malaysia, Iran and Bangladesh; and use of Claude in Yemen, China and Russia to develop software for conventional weapons, including firearms, missiles, armed drones, bombs and other munitions.
Comprehension
- Who does Anthropic say tried to misuse its AI models?
- What did some researchers do to get access for biology work?
- Why did Anthropic avoid naming institutions or countries?
- Name two places linked to surveillance or propaganda in the report.
- How many case studies focused on biological research?
Grammar focus
Ellipsis and substitution
Ellipsis and substitution let you avoid repeating words by replacing them with short forms (substitutes). Use words like do/does/did so, the same, or ones to replace a verb phrase or noun phrase. (Examples constructed.)
- Some researchers refused to publish, while others **did the same**.
- One lab used the model; **so did they** in private.
Grammar exercise
Circle the correct option in each sentence.
- Scientists published papers, and other teams (did them / did so) immediately.
- Anthropic reported misuse; industry groups may (do likewise / do like that).
- Some researchers raised alarms; others (so / too) stayed silent.
- A lab developed a tool, but others built better (one / ones).
- One study criticised policy; several others (did so / did likewise).
Discussion
- Should companies publish detailed misuse reports or keep them private?
- How strong should safeguards be when research could also help medicine?
- What duties do AI labs have if states use models for espionage?
- Which is more urgent: near-term misuse or far-future AI risks?
Vocabulary, reading, and exercises with instant feedback