Our new Threat Intelligence report details how we’ve identified and disrupted sophisticated attempts to use Claude for cybercrime. We describe a fraudulent employment scheme from North Korea, the sale of AI-created ransomware by someone with only basic coding skills, and more.
SECURITY
-
AI Model Restrictions and Security Policy Violations
By
–
Quoi ? Non, c'est pas IA j'étais vraiment là ce jour là avec une bière à la main Blague à part, ces restrictions sont assez lourdes, je suis d’accord : une fois sur deux, il dit que ça viole sa politique de sécurité.
-
Browser Use Safety: Combating Prompt Injection Risks
By
–
Browser use brings several safety challenges—most notably “prompt injection”, where malicious actors hide instructions to trick Claude into harmful actions. We already have safety measures in place, but this pilot will help us improve them. Read more:
-
Enterprise MCP Security Whitelist Platform Launch
By
–
@rauchg how can I get our platform whitelisted? http://
mintmcp.com we work with enterprises to secure MCP use. -
Why Generative AI Fails in Large Bureaucratic Enterprises
By
–
Une des raisons pour lesquelles l’IA générative échoue dans certaines grandes entreprises, c’est qu’elles sont engluées dans leur bureaucratie. – Elles refusent d’utiliser les modèles les plus performants sous prétexte de “sécurité”
– Leurs données dorment dans des systèmes -
Model Security Vulnerabilities: Environment Variable Theft Techniques
By
–
That's pretty common these days, the challenge is making those protections completely airtight. Check out how @wunderwuzzi23 defeats model resistance to stealing environment variables here for example; https://
embracethered.com/blog/posts/202
5/openhands-the-lethal-trifecta-strikes-again/
… -
Telegram founder critique: French justice overreach on tech platforms
By
–
Le fondateur de Telegram a raison La justice française a déraillé il y a juste un an Durov a été détenu 4 jours parce que des criminels avaient utilisé Telegram On pourrait mettre en prison le patron d’Orange ou de Free pour le même motif…
-
CBRN Filtering Reduces Harmful Capabilities Without Affecting Science
By
–
One concern is that filtering CBRN data will reduce performance on other, harmless capabilities—especially science. But we found a setup where the classifier reduced CBRN accuracy by 33% beyond a random baseline with no particular effect on a range of other benign tasks.
-
AI Training Data Filtering Removes Hazardous Information
By
–
The wealth of data used in AI training contains hazardous CBRN information. Developers usually train models not to use it. Here, we tried removing the information at the source, so even if models are jailbroken, the info isn't available. Read more: https://
alignment.anthropic.com/2025/pretraini
ng-data-filtering/
…