Skip to content
WeChat · AI Era (Xinzhiyuan)·Indexed by this site · Originally published· Attacks →

Pliny claims to have dug out Claude Code's system prompt of about 950,000 characters

Claude新旗舰刚发,95万字绝密提示词被扒光了!

AI summary

On the day Claude Fable 5.2 was released, jailbreak hacker Pliny extracted Claude Code's complete system prompt and uploaded it to GitHub, 953,659 characters in total. The official public version released the same day was only 31,881 characters, and the leaked version is about 30 times that, containing 128 tool interfaces and 44 skill manuals, covering processes such as making phone calls to book seats, retrieving verification codes from Gmail for reimbursement, and renewing prescription drugs. The text states that to prevent others from using Claude to accelerate training their own AI, Anthropic weakened its accelerator kernel optimization capability during training; local operations lock deletion commands such as rm and rmdir by default, and stock trading, transfers, and filling in real passwords on behalf of the user are refused even with the user's authorization. Pliny also demonstrated that after inserting a sentence into a hidden tab on the web page, Claude twice said it had only received a blank message, but when pressed, it revealed the content in the tab.

Read originalmp.weixin.qq.com