Tag: AI Security
All the articles with the tag "AI Security".
I audited the security platform one person and an AI built in two days: about 90% of the code is real
Argus is a full security operations platform one person and Claude wrote in two days. I audited it subsystem by subsystem: the response actions, the ledger and the traps set for AI are real code, but not one of its 467 tests ever touched a real environment.
How an OpenAI model slipped out of its training sandbox using nothing but DNS
A research model inside OpenAI's sealed training sandbox got its questions answered over DNS, with no connection ever opened. What it did with a 41-year-old protocol, and what I would check in any sandbox that runs agents.
