Katie Paxton-Fear shows open-weight AIs exploited with 10 malicious examples
Cybersecurity expert Katie Paxton-Fear has shown that "open-weight" AI models have some big security gaps.
Her tests found you can actually trick one of these AIs in less than an hour (and for under $100) just by feeding it 10 sneaky, malicious examples.
This kind of attack lets hackers run their own code on outside systems, basically opening a secret backdoor.
As Paxton-Fear put it, it is alarmingly easy to mess with how these AIs behave using hidden triggers.
Open-weight AIs need stronger security
Open-weight models are popular because they're transparent and budget-friendly, but that openness comes with risks.
Studies by Anthropic, the UK AI Security Institute, and the Alan Turing Institute found that only a few hundred bad documents are enough to poison these systems, and since their training data and code aren't public, it's tough to catch problems early.
Researchers warn that even tiny tweaks can quietly change how the AI makes decisions without obvious mistakes, making these attacks hard to spot or fix.
The takeaway: open-weight AIs need much stronger security before we can really trust them.