This report is based on a paper presented to the USENIX Security Symposium. (Ironically, the web-site is currently unstable. Try hitting refresh.)
The symposium does not appear to do a full peer review of presentations, but has refereed papers and invited talks.
It isn't clear that the referees would be experts in neuroscience.
Therefore, it would be nice to see this novel work reproduced before accepting it.
The paper itself:
describes their experiment (with 28 students), where they had the subject calibrate the EEG device, and then attempted to detect when a known image was displayed to the subject compared to an unknown image. The subjects were prepped to think about particular subjects (first digit of their (temporary) PIN, where they lived, their month of birth, their bank and the people they knew).
They didn't actively "read" the actual information from the brain - they tried to measure whether parts of the brain were active when it was thinking about a certain image. Nor did they deliberately focus on particular signals, but trained on the input during calibration to find whatever signal was best at indicating recognition.
The correct answer was found by the first guess in
20% of the cases for the experiment with the PIN, the
debit cards, people, and the ATM machine. The location
was exactly guessed for 30% of users, month of birth for
almost 60% and the bank based on the ATM machines
for almost 30%.
So, a random guess of the first digit would be 10%, they brought that up to 20%. Further, they had a good second guess if the first one was incorrect. They measured their success in terms of the reduction in entropy, which for some measures ranged from 15% to 40%.
While they claim "the classifiers perform significantly better than the random attack", I didn't see an analysis of the significance, which seems a serious shortfall.
The paper performs a basic review of similar experiments.
In conclusion, the experimenters have come up with an interesting idea and produced a exploratory proof-of-concept paper.
However, this will need to be reproduced before it can be accepted. It will require a lot of new development and sophistication before it can be considered a serious threat and/or technology.