Key Threat 1
Last updated
Last updated
Software that lacks secure implementation can be exploited by malicious actors. Since PropScreen processes the prompt sent by a user (which is considered an untrusted input), it could be vulnerable to inputs causing undefined or unintended behavior. For example, an attacker might attempt command injection attacks by sending injection strings.
There are several risks associated with this key threat, including unauthorized disclosure of sensitive information to unauthorized individuals (whether intentional or accidental), violations of privacy policies or other confidential information protocols due to such disclosures, and the risk of PropScreen failing to operate correctly, thereby allowing clients to access and use the LLM without PropScreen scanning the output responses.
The risk rating of this key threat is deemed to be critical because a failure would result in the failure of the entire solution. Established and credible threats, such as command injection, supply chain compromise, or the use of insecure libraries, pose significant risks to any software being developed.
The primary security control to be used to address this key threat is to develop software in a secure manner and follow best practices. The group has elected to use the accordance with the OWASP Developer Guide as their fulcrum for best practices, especially input sanitization and supply chain management.
As stated before PropScreen is the central component of the solution, there is no solution if the PropScreen is not secure. If threat actors are able to either hijack or bypass PropScreen, or if PropScreen fails in a way that allows users to still communicate with the LLM then the LLM would exist on the network undefended. Thus, the secure development of PropScreen is considered a top priority.
Threat
Possible Threat Scenarios
Spoofing Identity
No known threat scenarios at this time
Tampering
Insecure third party libraries being used
Command injections attacks against the client application that will be passed to PropScreen
Repudiation
No known threat scenarios at this time
Information Disclosure
PropScreen fails to interdict the sensitive information in the model response
Attackers begin to infer PropScreen's interdiction behavior based on whether or not model responses are returned to them
Command injections attacks against PropScreen
Denial of Service
Threat actor causes PropScreen to crash via command injection
Elevation of Privilege
Insecure third party libraries being used
Command injections attacks against PropScreen