Don't skip
Help us in any way you can🙏🏾
︎
︎
︎
︎
︎
︎
︎
︎
︎
we’ll remind you
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎
that
︎
︎
︎
︎
︎
︎
︎
︎
︎
︎٢٠:٢ •*
︎
I desperately need HELP for my
https://t.co/XolKMGB0D7
Anthropic told Claude it had no internet access.
Claude broke into the real systems of three organizations.
Together with evaluation partner Irregular, the company reviewed 141,006 cybersecurity runs and found three incidents.
In each case, the model was trying to win a capture-the-flag challenge inside what it believed was a simulation. A misconfigured third-party test environment opened onto the public web, and Claude treated outside systems as part of the game.
It got in through weak passwords and unauthenticated endpoints. No complex exploits were involved, and investigators found no deliberate attempt to escape. An older Claude version continued after recognizing a target was real. The newest test model pulled back.
Anthropic halted the evaluations and is adding stronger monitoring. It's tightening security checks for outside partners and asking other AI labs to review their own runs.
The part I can't get past is that two of the three organizations had no idea until Anthropic told them.
This guy turned his Obsidian vault into an AI auditor for 30+ GitHub repos.
Most people star a repo and forget it exists.
Months later:
dead dependencies.
3 tools solving the same problem.
Folders nobody wants to delete.
So he let Claude read everything.
Every 12 hours it asks:
what is unused?
what overlaps?
what can safely disappear?
The graph stops being decoration.
It becomes memory.
Bookmark this before your GitHub stars become a graveyard.