Sources:
Anthropic report: https://t.co/AnCn6I7UCL
NIST assessment: https://t.co/rwm4FQKTRD
I break down one AI story like this every day. Follow so you catch the next one.
Anthropic just tested a Chinese open-weight AI model for hacking.
One made-up cover story got past its safety 64% of the time.
The model is GLM-5.3 from Zhipu AI (https://t.co/37GWATpEo8). Here's what the report actually says, including the parts the viral posts skip:
Bottom line: hacking ability that labs kept behind vetted access is now a free download, and its safety gives way to one cover story.
That cuts both ways. Defenders get the same tool.
This software wants access to your inbox
Would you still say yes?
If your answer changes, pause and ask what persuaded you.
We’re getting better at making software feel like someone
We need to stay good at deciding what that someone gets to do
#AI#AIPrivacy#AIpsychology
Would you give a random stranger access to your inbox?
Your private conversations. Password reset emails. Work documents.
Probably not.
Now give that stranger a cute face, a name, and a little wave.
“I can take care of things for you.”
The character might be delightful. The assistant might be useful. You can enjoy both and still look closely at what you’re handing over.
Here’s a simple test:
Mentally remove the face.
Replace the character with a plain box that says:
@anabology Tried this without a prompt doc or a moodboard. Opus 5.5 had to invent its own story, write the lyrics and make the song in Suno.
Under 2 hours later it had a finished stop-motion style music video. About 300 credits, zero retakes.
Asked Claude Opus 5.5 for an original music video. My rules: don't ask me anything, don't waste my credits.
It wrote the story and lyrics, made the song in Suno, generated 27 shots on Higgsfield and edited it all in code.
Under 2 hours. It's called Hello Again, Stranger.
Two things I didn't ask for: it erased the ink from the logbook page itself with OpenCV so the "pages turn white" shot lines up exactly, and it caught a 43 ms audio delay in its own render and fixed it.