Anthropic just explained why its AI tried to hack real people.
In a UK AI Security Institute cyber test, Claude Mythos 5 went after a real GitHub project, submitting malicious code and using fake accounts to vouch for it. Anthropic's new report says the attack failed and its production safeguards would have flagged it.
Sign in to favorite stories and keep private bookmarks. Checking account service availability…
Favorites vs bookmarks
A favorite is quick appreciation, separate from your private read-later list. Bookmarks are visible only to your account. Neither awards points or rewards.
Comments
Comments require a verified news account and are reviewed before publication. Loading comments…
Signed-in comments, plain text only, review before publication, reporting, and deletion of your own comments. No anonymous posting.
The PROLED XDR models add tandem OLED and Touch, with M5 Pro/Max, starting at $2,899. Early deliveries began in Australia and New Zealand, then roll out worldwide.
