There are also tons of ways to exfiltrate data through known channels in ways that are difficult for security researchers to distinguish from otherwise secure app analytics code.
A crash/exception logging system, say, might appear to researchers to anonymize data, but it would be very possible for code to be written that happens to raise a mundane exception when specific users or geofences see specific words on screen, in a way where that list of users/geofences/words could be controlled by non-technical teams. The log message itself doesn't even need to carry sensitive data; its existence alone, when the trigger conditions are known, can be used to carry out a highly targeted attack.
Even open-source systems can be vulnerable to this: see e.g. https://github.com/signalapp/Signal-iOS/blob/eaed4da06347a3a... and consider the ways it might be possible for a small group of people at Signal to cause a specific set of messages to be seen as corrupt without raising any flags to the community auditing the code.
Of course, lack of visibility into runtime errors can lead to vulnerabilities as well. I don't think the solution is for us as a community to advocate for removing all error analytics in distributed systems. But we can't ever forget that: all analytics surfaces are attack surfaces.
A crash/exception logging system, say, might appear to researchers to anonymize data, but it would be very possible for code to be written that happens to raise a mundane exception when specific users or geofences see specific words on screen, in a way where that list of users/geofences/words could be controlled by non-technical teams. The log message itself doesn't even need to carry sensitive data; its existence alone, when the trigger conditions are known, can be used to carry out a highly targeted attack.
Even open-source systems can be vulnerable to this: see e.g. https://github.com/signalapp/Signal-iOS/blob/eaed4da06347a3a... and consider the ways it might be possible for a small group of people at Signal to cause a specific set of messages to be seen as corrupt without raising any flags to the community auditing the code.
Of course, lack of visibility into runtime errors can lead to vulnerabilities as well. I don't think the solution is for us as a community to advocate for removing all error analytics in distributed systems. But we can't ever forget that: all analytics surfaces are attack surfaces.