|
To reproduce this issue, a policy must be applied to ChatGPT traffic with a DLP security profile and SSL deep inspection enabled. The expected behavior is that the system filters posts containing credit card numbers by leveraging the built-in credit card data type to match the traffic, including cases where the credit card numbers are presented as a list separated by new lines.
Steps to reproduce:
- A DLP policy applied to ChatGPT traffic is needed. This policy must have SSL deep inspection enabled and the dlp profile applied:
- A DLP security profile configuration with a sensor configured with a dictionary matching DLP built in datatype:
- A ChatGPT post containing credit card numbers separated with new lines:
- Notice that the post was not blocked by DLP, even though it contains multiple valid credit card numbers.
Root Cause:
Upon analysis, it was found that this behavior is related to how ChatGPT formats data before sending it to the server.
- ChatGPT submits the content as JSON payloads.
- In decrypted HTTP POST requests, the data appears as a JSON string formatted this way, with the \n separator embedded in the string:
are these credit card numbers?\n4111111111111111\n4111 1111 1111 1111\n5555555555554444\n....
- The built-in Credit Card data type uses this regular expression with word boundaries (\b):
\b([2-6]{1}\d{3})[- ]?(\d{4})[- ]?(\d{2})[- ]?(\d{2})[- ]?(\d{2,4})\b
- Since the newline character \n is embedded into the string, it does not constitute a word boundary, so the regex fails to match credit card numbers.
Solution / Workaround:
To address this, a custom dictionary data type can be created for credit card detection without using the \b boundary included in the built-in data type.
This ensures that credit card numbers within ChatGPT traffic are detected correctly, even when separated by new lines.
Example configuration with the fix:
With this configuration, traffic is matched by DLP sensor and it is blocked correctly:
CLI Configuration with solution applied:
config firewall policy
edit 24
set name "chatgpt"
set srcintf "fortilab-local-zone"
set dstintf "virtual-wan-link"
set action accept
set srcaddr "all"
set dstaddr "chatgpt"
set schedule "always"
set service "ALL"
set utm-status enable
set inspection-mode proxy
set ssl-ssh-profile "deep-inspection"
set dlp-profile "block-cc-profile"
set application-list "default"
set logtraffic all
set nat enable
next
end
config dlp profile
edit "block-cc-profile"
set feature-set proxy
config rule
edit 1
set name "block-cc"
set type message
set proto http-post
set filter-by sensor
set sensor "cc"
set action block
next
end
next
end
config dlp sensor
edit "cc"
config entries
edit 1
set dictionary "custom-cc"
next
end
next
end
config dlp dictionary
edit "custom-cc"
config entries
edit 1
set type "regex"
set pattern "/([2-6]{1}\\d{3})[- ]?(\\d{4})[- ]?(\\d{2})[- ]?(\\d{2})[- ]?(\\d{2,4})/m"
set repeat enable
next
end
next
end
Important considerations:
- The modified regex is less restrictive than the built-in Credit Card data type.
- This may result in false positives in some environments.
- This workaround is intended specifically for ChatGPT transactions. Other applications or websites that format strings differently may not require this adjustment.
|
